Abstract
The potential of technology to revolutionize healthcare is exemplified by the synergy between artificial intelligence (AI) and early detection of cardiomegaly, demonstrating the power of proactive intervention in cardiovascular health. This paper presents an innovative approach that leverages advanced AI algorithms, specifically deep learning (DL) technology, for the early detection of cardiomegaly. The methodology consists of five key steps, including data collection, image preprocessing, data augmentation, feature extraction, and classification. Utilizing chest X-ray (CXR) images from the National Institutes of Health (NIH), the study applies rigorous image preprocessing operations, including color transformation and normalization. To enhance model generalization, data augmentation is employed, paving the way for two distinct DL models, a convolutional neural network (CNN) developed from scratch and a pretrained residual network with 50 layers (ResNet50), and adapted to the problem domain. Both models are systematically evaluated with five optimizers, revealing the AdaMax optimizer’s superiority for the CNN model and AdaGrad’s efficacy for the modified ResNet50. The proposed CNN with AdaMax achieves an impressive 99.91% accuracy, outperforming recent techniques in precision, recall, and . This research underscores the transformative potential of AI in cardiovascular health diagnostics, emphasizing the significance of timely intervention.
1. Introduction
Cardiomegaly, or an enlarged heart, may be a sign of cardiac insufficiency in a variety of medical conditions, including high blood pressure, coronary artery disease, heart valve disease, and pulmonary hypertension. Detecting cardiomegaly early on is of utmost importance, and one of the most widely used noninvasive and affordable medical imaging tests for early diagnosis is chest X-ray (CXR) imaging [1]. Recent advances in deep learning (DL), combined with the availability of comprehensive CXR databases, have greatly improved the performance of computer-aided detection for cardiomegaly. These methods have achieved results that are comparable to those of human radiologists [2–6]. However, it is important to note that current DL-based detection methods rely heavily on binary classification of the entire CXR, using image-level label-dependent learning to detect cardiomegaly. Although classification-based methods have limitations due to the unclear process by which DL algorithms arrive at conclusions, segmentation-based methods can effectively identify the boundaries of the lungs and heart on CXRs, enabling the automatic calculation of the cardiothoracic ratio (CTR). This index is highly valuable for evaluating cardiac enlargement. It is essential to possess a thorough comprehension of the standard ranges for the CTR since these ranges are subject to variation due to factors such as age, gender, and population demographics. By comparing the calculated CTR value to established norms, we can identify possible indications of cardiac enlargement. An elevated CTR value signals an imbalance between heart size and thoracic dimensions, which can be a sign of cardiomegaly. The CTR is determined by dividing the maximum horizontal cardiac diameter by the maximum horizontal thoracic diameter. The normal range for CTR is between 0.42 and 0.50. If the value is higher than 0.50, it indicates cardiomegaly [7]. To accomplish this, it is important to identify key anatomical landmarks in the chest’s posterior-anterior (PA) projection. Calibrated measurement tools are used to determine the maximal horizontal diameter of the cardiac silhouette, which includes the outer boundaries of the heart, such as the left ventricular border and the right atrial margin. At the same time, the thoracic cavity’s widest transverse diameter is measured, establishing the dimensions of the anatomical context in which the heart resides. The quotient of these two measurements, the cardiac diameter and the thoracic width, gives rise to the CTR.
Interpreting the CTR requires a nuanced understanding of normative ranges, which vary based on many factors such as age, gender, and population demographics. Comparing the calculated CTR against established norms is crucial in detecting cardiac enlargement. Elevated CTR values indicate an altered equilibrium between heart size and thoracic dimensions, which can signify cardiomegaly. To calculate the CTR, you need to follow these key steps:(1)Identify Cardiac Borders. Discern the outer boundaries of the cardiac silhouette on the CXR image. This involves identifying anatomical landmarks, such as the left ventricular border on the left side and the right atrial border on the right side.(2)Measure Cardiac Width. Using calibrated measurement tools, ascertain the widest horizontal diameter of the cardiac silhouette. This encompasses the distance between the identified borders, encapsulating the entire extent of the heart’s dimensions visible on the CXR.(3)Measure Thoracic Width. Simultaneously, measure the widest transverse diameter of the thoracic cage. This measurement corresponds to the broadest dimension of the chest cavity as depicted on the CXR.(4)Calculate the CTR. Divide the measured cardiac width by the measured thoracic width to obtain the CTR value. Mathematically, it is expressed as
The CTR value is a unitless ratio that quantifies the heart’s size relative to the thoracic cavity. An elevated CTR value suggests cardiomegaly. However, the quantitative datum is not considered in isolation. Its clinical significance is accentuated through contextual integration with patient-specific variables, including symptoms, medical history, and associated diagnostic data. A thorough analysis of an elevated CTR can prompt further clinical exploration, including specialized consultations, corroborative diagnostic investigations, and targeted therapeutic interventions.
In summary, the CTR from CXR is a multidimensional diagnostic tool that provides early insights into cardiomegaly. By embracing quantitative and qualitative aspects of this approach, healthcare practitioners gain a nuanced framework for identifying cardiac enlargement, leading to proactive clinical management and better patient outcomes.
DL is a powerful technique [8] that involves training artificial neural networks to analyse data and make predictions. Its ability to identify complex patterns and accurately predict outcomes based on large datasets has made it increasingly popular across various fields. For instance, DL can be utilized to detect cardiomegaly (enlarged heart) in CXR images, offering a practical solution to diagnose this condition. The importance of DL in the early detection of cardiomegaly through CXR images lies in the following:(1)Feature Extraction. DL models can automatically learn relevant features from CXR images without the need for explicit feature engineering. This enables the model to capture intricate details that might be indicative of cardiomegaly.(2)Complex Pattern Recognition. Cardiomegaly detection involves recognizing subtle variations in heart size, shape, and contour. DL models excel at identifying complex and nonlinear patterns within the data, which may not be easily discernible by traditional methods.(3)Scalability. DL models can be trained on large datasets, allowing them to generalize from a diverse range of CXR images. This scalability enables better performance in identifying various manifestations of cardiomegaly across different patient populations.(4)Continuous Learning. DL models can be fine-tuned and updated as new data become available. This adaptability ensures that the model remains relevant and accurate over time, improving its effectiveness in detecting cardiomegaly.(5)Reduction of Human Error. Automated analysis of CXR images using DL can minimize human subjectivity and errors that may arise from manual interpretation, leading to more consistent and reliable results.(6)Early Intervention. By accurately detecting signs of cardiomegaly in its early stages, healthcare professionals can initiate appropriate interventions, such as further diagnostic tests or treatment plans, to mitigate the progression of cardiovascular diseases.
DL technology for early detection of cardiomegaly involves constructing neural networks with multiple layers that can learn to map CXR images to relevant diagnostic outcomes. These networks are trained using labeled datasets, where CXR images are paired with corresponding labels indicating the presence or absence of cardiomegaly. The learning process involves optimizing the model’s parameters to minimize the difference between predicted outcomes and actual labels, thereby enabling the model to generalize to new, unseen CXR images. Overall, the application of DL technology in the early detection of cardiomegaly through CXR images offers a sophisticated and data-driven approach to improve diagnostic accuracy, potentially leading to enhanced patient outcomes and better management of cardiovascular health.
1.1. Motivations
The primary objective of this paper is to develop a DL model for early identification of cardiomegaly. The paper proposes two distinct DL models to solve a given problem. The first model is a convolutional neural network (CNN) which has been developed from scratch. The second model is based on a pretrained residual network with 50 layers (ResNet50) which has been modified to make it more appropriate for the problem at hand. To evaluate the proposed models, the research considers five types of optimizers, namely, Adaptive Gradient (AdaGrad), Adaptive Moment Estimation (Adam), Adaptive Moment Estimation with Infinity Norm (AdaMax), Nesterov-accelerated Adaptive Moment Estimation (NAdam), and Root Mean Square Propagation (RMSprop). The effects of each optimizer on the performance, in terms of accuracy, precision, recall, F1-score, and processing time, of the models are analysed and discussed in detail.
The proposed model comprises seven distinct steps. Initially, a relevant dataset for cardiomegaly is gathered. After that, the unstructured data are preprocessed, converting them into a structured format for classification purposes. Next, the data are split into training and testing sets, with data augmentation being employed to improve the model’s predictive abilities. Simultaneously, DL models are utilized for both feature extraction and classification. Finally, each DL model is compiled with five types of optimizers, and their effectiveness is measured using predefined metrics. Based on the evaluation results, the best optimizer for each DL model is determined, ultimately leading to the identification of the best-performing DL model for early cardiomegaly detection.
1.2. Contributions
This paper introduces a pioneering methodology that incorporates both deep learning model diversity and optimization strategy comprehensiveness. The standout features and contributions, particularly considering the noteworthy results obtained, are outlined as follows.
1.2.1. Architectural Innovation
(1)The paper goes beyond conventional methods by employing two distinct DL models—a custom-built CNN and a modified ResNet50.(2)This dual-model paradigm allows for a nuanced comparison, showcasing the ingenuity of a model created from scratch against the adaptability of a modified pretrained model. The exceptional results affirm the efficacy of both approaches in early cardiomegaly detection.1.2.2. Optimizer Diversity and Performance Metrics
(1)The study systematically evaluates both models under the influence of five diverse optimization algorithms namely, AdaGrad, Adam, AdaMax, NAdam, and RMSprop.(2)The comprehensive exploration of optimization strategies is coupled with a multifaceted evaluation using metrics like accuracy, precision, recall, and F1-score. The exceptional achievements, such as 99.91% accuracy with AdaMax for the proposed CNN and 99.73% accuracy with AdaGrad for the modified ResNet50, underscore the significance of this dual exploration.1.2.3. Efficiency as a Decisive Criterion
(1)The paper introduces computational efficiency as a crucial criterion, evaluating the speed of predictions for each model-optimizer combination.(2)Recognizing AdaMax as the fastest optimizer for the proposed CNN and AdaGrad for the modified ResNet50 not only emphasizes the practical implications for real-time clinical scenarios but also highlights the efficiency gains achieved through strategic optimizer selection.1.2.4. Optimal Configurations and Model Superiority Determination
(1)The study not only identifies optimal configurations for each model-optimizer pair but also determines the superior model-optimizer combination for both the proposed CNN and the modified ResNet50.(2)The exceptional results, such as 99.91% accuracy, 100% precision, 99.40% recall, and 99.7% F1-score with AdaMax for the proposed CNN, and 99.73% accuracy, 98.8% recall, and 99.1% F1-score with AdaGrad for the modified ResNet50, reinforce the practical applicability and superiority of these configurations.So, these findings establish the study as a pivotal contribution toward improving the effectiveness and practicality of DL in medical image analysis.
1.3. Structure
The rest of the paper is structured into various sections. Initially, the prior research is presented in Section 2. After that, Section 3 discusses a proposed system for early detection of cardiomegaly, which is further divided into two subsections: Section 3.1 presents suggested DL models, while Section 3.2 provides suggested optimizers for each DL model. Proceeding further, Section 4 presents the analysis and experimental findings, which are then compared to some state-of-the-art methods. Finally, the paper concludes in Section 5 by summarizing the findings and suggesting further research.
2. Literature Review
The pursuit of precise and timely identification of cardiomegaly, an imperative indicator of underlying cardiovascular pathologies, is a paramount objective in medical imaging. Within this context, the examination of chest radiographs, or CXRs, emerges as a pivotal diagnostic modality. However, the inherently intricate and context-dependent nature of cardiomegaly poses challenges to traditional interpretative paradigms. To address this, a surge of interest has been directed toward harnessing the capabilities of DL methodologies, particularly CNNs, to facilitate the automated detection of cardiomegaly from CXR images. This section embarks on an incisive exploration of the scholarly landscape, traversing methodologies underpinning the application of CNNs in this domain as in Table 1. The investigation encompasses the diverse spectra of data augmentation techniques, architectural configurations, and training modalities that underlie the DL paradigms employed in tackling the complexity of cardiomegaly detection. Additionally, an analytical examination of reported evaluation metrics, such as precision-recall curves and F1-scores, resonates with the overarching goal of elucidating the advancements and potential limitations of CNN-based strategies in enabling robust and clinically relevant identification of cardiomegaly. The synthesis of these endeavors forms a crucial substrate, illuminating the trajectory of employing DL, specifically in the realm of CXR-based cardiomegaly detection, thus paving the way for a more comprehensive understanding of the scientific landscape and its implications. Lin et al. [9] introduced an innovative classifier for automated cardiomegaly-level screening in frontal PA view CXR images. The multilayer one-dimensional (1D) CNN employed dual-round 1D convolutional processes, enhancing the recognition of normal conditions and different cardiomegaly levels. The classification layer utilized a grey relational analysis-based classifier for computational simplicity. Evaluation through 10-fold cross-validation demonstrated promising results with precision of 97.80%, recall of 98.20%, accuracy of 98.00%, and an impressive F1-score of 97.99%. The study utilized datasets from the National Institutes of Health (NIH) CXR image collection, showcasing the potential of advanced image analysis in medical diagnostics. Wu et al. [10] developed a hybrid approach employing both 2D (two-dimensional) and 1D CNN-based classifiers for swift cardiomegaly screening in clinical applications, specifically utilizing frontal PA CXR scans. The methodology integrates 2D and 1D convolutional processes, along with a multilayer interconnected classification network, to enhance feature extraction and pattern recognition tasks. This holistic strategy aims to enrich the original CXR images while eliminating undesired artifacts. The classifier is trained and assessed using 10-fold cross-validation. The experimental results exhibit promising outcomes aligned with the intended medical purpose, boasting a precision of 97.60%, recall of 99.20%, accuracy of 98.40%, and an impressive F1-score of 98.38%. Notably, datasets from the NIH CXR image collection were utilized.
Chen et al. [11] presented a novel approach for cardiomegaly detection on CXR images, introducing a dual attention network named CXRDANet. The CXRDANet incorporates a channel attention module (CAM) and a spatial attention module (SAM), strategically enhancing features associated with the lesion region. Experimental results showcase the effectiveness of their technique, achieving an accuracy of 90.50%, a recall of 94.45%, and an F1-score of 90.59%. The training and test sets involve nonoverlapping CXR images of cardiomegaly and normal cases sourced from CXR14 and NLM-CXR within the NIH CXR image collection. Zhou et al. [12] employed a transfer learning approach in their study to identify cardiomegaly in CXR images, as detailed in their work. The researchers achieved superior results by utilizing a combination of InceptionV3 and ResNet50 pretrained models, yielding impressive AUCs of 86.0%, surpassing the outcomes of previous similar studies. Their research highlights the effectiveness of employing transfer learning methodologies to develop a CXR computer-aided detection (CAD) system. The proposed technique exhibits an accuracy of 79.7% and an F1-score of 80.00%, validated on the CXR8 dataset sourced from the NIH CXR image collection. Innat et al. [13] introduced Cardio-XAttentionNet, a DL network designed for accurate cardiomegaly classification and localization in CXRs. The model incorporates a convolutional attention mapping technique, enhancing the global average pooling (GAP) system through the addition of a weighting term to form an efficient attention mapping mechanism (AMM). This innovation enables the detection of cardiomegaly through both image-level classification and pixel-level localization, even with image-level labeling alone. The proposed attention mapping network’s backbone integrates robust ConvNet designs. The best-performing single model demonstrates impressive overall accuracy, recall, F-1 measure, and area under the curve (AUC) scores of 87.00%, 85.00%, 86.00%, and 89.00%, respectively. Training of the model utilized the CXR14 dataset from the NIH.
Ajmera et al. [14] presented a DL-based algorithm aimed at automating CTR computation, facilitating swift cardiomegaly diagnosis and enhancing radiological workflow. The algorithm employs the Attention U-Net DL architecture for automatic CTR calculation. An observer performance test was conducted to assess radiologists’ efficacy in identifying cardiomegaly, both with and without the assistance of artificial intelligence. The U-Net model exhibits commendable performance metrics, with a recall of 80.00%, precision reaching 99.00%, and an F1-score of 88.00%. These results underscore the potential of the DL algorithm in aiding radiologists in the efficient and accurate diagnosis of cardiomegaly. Sarpotdar et al. [15] introduced a DL-based, customized, and retrained U-Net model designed for the diagnosis of cardiomegaly. Their model integrates data preprocessing, image enhancement, image compression, and classification processes to optimize computation time. Utilizing a CXR image dataset for simulation, the study achieves notable diagnostic metrics, including an accuracy of 94%, recall of 96.2%, and specificity of 92.5%. The evaluation was based on the CXR8 dataset sourced from the NIH CXR image collection. Candemir et al. [2] proposed a method for cardiomegaly detection in CXRs, as outlined in their work. The algorithm comprises two key stages: identification of heart and lung regions in CXRs and extraction of radiographic index from their boundaries. A lung detection algorithm, extended for automatic determination of heart boundaries, was employed to locate these regions. Standard models of heart and lung regions, learned from a public CXR dataset with boundary markings, were registered to patient CXR images to estimate region locations. Traditional and recently published radiographic indexes were utilized for index computation. The classifier combined these indexes, successfully identifying patients with cardiomegaly with an accuracy of 77.00% and sensitivity (recall) of 77.00%. The evaluation was conducted on the JSRT Set (compiled by the Japanese Society of Radiological Technology), comprising 247 CXRs, and the Indiana Set, consisting of approximately 4,000 frontal and lateral CXRs with various lung abnormalities. From the latter, 250 frontal CXRs with cardiomegaly and 250 normal cases were selected for testing. In a recent investigation conducted by Bougias et al. [16], the analysis of CXR aimed to discern cases of cardiomegaly. The study utilized a dataset comprising 2000 CXRs, evenly divided between normal and confirmed cardiomegaly cases. Several DL networks, including Google’s InceptionV3, VGG16, VGG19, and SqueezeNet, were employed to extract deep features from the CXRs. Subsequently, these extracted features were utilized for classifying CXRs as either indicative or not indicative of cardiomegaly through a logistic regression algorithm. Among the DL networks, VGG19 demonstrated the most favorable results, achieving an overall accuracy of 84.5%. The other networks exhibited varying accuracy ranges, with sensitivities ranging from 64.1% to 82%, specificities ranging from 77.1% to 81.1%, positive predictive values (PPVs) ranging from 74% to 81.4%, and negative predictive values (NPVs) ranging from 68% to 82%.
Ribeiro et al. [17] introduced a DL model designed to detect cardiomegaly using CXR images, accompanied by an evaluation of three local explainable methods: Grad-CAM, LIME, and SHAP. The study’s findings reveal that their DL model achieved an accuracy of 91.8 ± 0.7%, precision of 74.0 ± 2.7%, sensitivity of 87.0 ± 5.5%, specificity of 92.9 ± 21.2%, F1-score of 79.8 ± 21.9%, and an area under the receiver operating characteristic curve (AUROC) of 90.0 ± 20.7%. The evaluation was conducted on the VinDr-CXR dataset, comprising CXR scans obtained retrospectively from two major hospitals in Vietnam. Table 1 summarizes the findings of the literature review.
Singh et al. [18] proposed using DL techniques to diagnose glaucoma, which is the second leading cause of permanent blindness. Their study showed that Inception-ResNet-v2 and Xception models outperformed other models in detecting glaucoma from fundus pictures. The automated system has great potential to improve early detection of glaucoma while reducing human efforts and time.
In a recent study by Khanna et al. [19], CT images and test kits were utilized to predict COVID-19 with success. By employing AI-powered chest X-ray analysis, automated analysis has been shown to greatly enhance accuracy and speed up the diagnostic process. The most dependable models, including the ensemble deep transfer learning CNN model and hybrid LSTM-CNN, have demonstrated the ability to deliver prompt and precise predictions, offering a meaningful advancement for clinical practice.
Khanna et al. [19] presented three innovative CNNs that can detect and classify diabetic retinopathy into five categories. The proposed models outperform many existing models in terms of classification, achieving a maximum accuracy of 0.9545, a maximum F1-score of 0.9685, a maximum sensitivity of 0.9566, and a maximum AUC score of 0.9769. The experiment results demonstrate that the proposed models are effective in identifying retinal issues in diabetic patients, leading to better diagnosis and preventing vision loss.
Table 1 introduces the state-of-the-art studies in the area of detecting cardiomegaly disease based on CXR images.
In discussing medical and disease data, it is crucial to prioritize accuracy and sensitivity since human life is at stake. Previous studies have shown that many of them need more balanced data, leading to subpar results in addressing medical problems. Our paper aims to address this issue by introducing changes to the same dataset used in these studies to balance the data as in Section 4.1 and achieve more precise and reliable outcomes. Our goal is to explore multiple DL models for early cardiomegaly detection using different optimizers to determine the best one for each model. By doing so, we aim to present a model that outperforms previous studies in the same field.
3. The Proposed Methodology
An early detection system for cardiomegaly is proposed in this paper, which utilizes specific DL models. The system involves several key steps, starting with the preprocessing of CXR images using a dedicated pipeline. This pipeline includes techniques such as data resizing, labeling, normalization, and color transformation. The preprocessed dataset is then divided into training and testing sets, and data augmentation is performed for both sets. The recommended DL models are finally constructed and trained based on these sets, and their performance is evaluated. The DL approach that has been suggested involves the use of two DL models, namely, CNNs and ResNet50. Additionally, five different optimizers: AdaGrad, Adam, NAdam, AdaMax, and RMSprop, will be utilized to observe their impact on each proposed model and its outcomes. Automated feature extraction and classification tasks are carried out by the DL models presented. The training accuracy and validation loss values are evaluated periodically. After that, the effectiveness of the available DL models is measured in terms of various assessment measures such as accuracy, precision, recall, F1-score, and processing time. The general design of the suggested system for early detection of cardiomegaly is illustrated in Figure 1.
3.1. Suggested Deep Learning Models
Two types of DL models, namely, CNNs and ResNet50 models, are recommended for use. These models are briefly explained in the following subsections. Additionally, each of the models is compiled with five different optimizers, namely, AdaGrad, Adam, NAdam, AdaMax, and RMSprop, which are also briefly explained in Section 3.2. The primary objective of using these models with these optimizers is to determine how each optimizer affects each proposed model and identify the best optimizer from the five that balances testing time and detection accuracy.
3.1.1. Convolutional Neural Network
Convolutional neural networks (CNNs) have brought about a revolution in the field of DL, with their incredible success in various domains such as image classification [20], object detection [21], and image segmentation [22]. As a result, there has been a significant increase in interest from both academia and industry in recent years. CNNs aim to make use of the spatial hierarchies in the data by utilizing convolutional and pooling layers, as explained in [23]. Convolutional layers are considered the key component of CNNs, which apply convolution operations to the input data through the use of a small filter, also known as a kernel that is positioned over the spatial dimensions of the input image. For a 2D input image and a 2D filter, the convolution operation can be mathematically expressed aswhere S(i, j) is the value at position (i, j) in the output feature map, I represents the input image, and K is the filter/kernel. The summation runs over all m and n values that correspond to the filter’s dimensions. After convolution, nonlinear activation functions like rectified linear unit (ReLU) are often applied elementwise to introduce nonlinearity into the model. Mathematically, the ReLU function is defined by the following equation:
Pooling layers, such as MaxPooling, are considered another crucial component of CNNs. Pooling layers help reduce the spatial dimensions of the feature maps while retaining important information. MaxPooling, for example, involves selecting the maximum value within a small region of the input. This operation is expressed as
The combination of convolutional layers and pooling layers helps CNNs capture and hierarchically represent features at different levels of abstraction. This ability is crucial for recognizing complex patterns and structures within images. In addition to convolutional and pooling layers, CNNs often contain fully connected layers near the end of the architecture. A fully connected layer, also known as a dense layer, is a fundamental building block in neural networks. It is a type of layer where each neuron or node is connected to every neuron in the previous layer, forming a fully connected graph. This layer is commonly used to capture complex relationships and patterns within the data, making it suitable for tasks such as classification and regression. Mathematically, let us consider a fully connected layer that takes input from a previous layer with n neurons. Each neuron in the fully connected layer applies a weighted sum of the input values followed by an activation function to produce its output. The output of the neurons in the fully connected layer can be calculated using the following equation:where is the weight connecting the neuron in the previous layer to the neuron in the fully connected layer, is the output of the neuron in the previous layer, is the bias term for the neuron in the fully connected layer, and the activation function introduces nonlinearity to the output.
In the context of neural networks, multiple fully connected layers can be stacked together, typically with activation functions in between, to form deep architectures. The deep neural network can learn increasingly complex and abstract representations of the input data as it goes through these layers. Fully connected layers have a significant number of parameters, often leading to a high computational load and potential overfitting. Regularization techniques like dropout and weight decay are commonly employed to mitigate overfitting. Additionally, global average pooling and convolutional layers have gained popularity, especially in convolutional neural networks, as they reduce the number of parameters and capture spatial hierarchies more effectively. A typical CNN architecture along with these layers is depicted in Figure 2. In this paper, the suggested CNN model involves several convolutional and pooling layers to discover the preprocessed images’ features and do the classification task. The following layers make up the recommended CNN model:(1)The first layer is a convolutional (Conv2D) layer with 32 filters of size 3 X 3, and incorporates the ReLU activation function. This is formally presented in equation (3), with the layer receiving an input tensor with dimensions of (224, 224, 3).(2)The second layer is a max-pooling (MaxPooling2D) layer, which has the maximum output from the neighborhood, with a pool size of 2 × 2. This layer reduces the spatial dimensions of the feature maps outputted by the convolutional layer, as explained in equation (5).(3)The third layer is another convolutional layer (Conv2D) with 64 filters and the ReLU activation function.(4)The fourth layer is another max-pooling (MaxPooling2D) layer with a pool size of 2 × 2.(5)The fifth layer is another convolutional layer (Conv2D) with 128 filters and the activation function.(6)The sixth layer is another max-pooling (MaxPooling2D) layer with a pool size of 2 × 2.(7)The seventh layer is another convolutional layer (Conv2D) with 128 filters and the activation function.(8)The eighth layer is another max-pooling (MaxPooling2D) layer with a pool size of 2 × 2.(9)The ninth layer is a flattened layer (Flatten) that converts the 2D feature maps outputted by the previous layer into a 1D vector.(10)The tenth layer is a dense layer (Dense). It is fully connected because all input neurons are considered by each output neuron, with 128 units and a activation function.(11)In the final layer, there is a dense layer that contains 2 units and utilizes the activation function. This function generates a probability distribution for the two classes of data. The function is proficient in transforming a vector of real values into a vector of real values that add up to 1. It can handle values that are positive, negative, zero, or greater than one and transforms them into values ranging from 0 to 1. The mathematical representation for is represented in the following equation:
In conclusion, following the comprehensive delineation of the 11-layer CNN architecture proposed for precocious detection of cardiomegaly, the pseudocode for the recommended CNN framework is delineated in Algorithm 1 and 2 that provide the pseudocode for the proposed CNN model and the modified ResNet50 model, respectively. In addition, the architecture of the suggested CNN model is illustrated in Figure 3, which provides a visual representation of a neural network’s architecture. It offers a structured and static view of how the various layers are connected and how data flow through the network. It illustrates the arrangement of input and output layers, hidden layers, and any branching or merging of connections. It captures the overall structure of the mode and gives a snapshot of the model’s design, making it suitable for conveying a high-level overview of the neural network’s layout and connections, and Figure 4 offers an interactive and dynamic representation of a neural network’s architecture. This interactive nature empowers users to explore intricate model configurations, understand data flow, and delve into specific layer properties, making it particularly beneficial for educational purposes and comprehending complex neural network architectures.
|
|
3.1.2. Residual Network with 50 layers (ResNet50)
Learning through residual mechanisms is a simplified process that focuses on the difference between the input and desired mapping rather than attempting to learn the entire mapping directly. This approach allows for effective learning, even in situations where the network is extremely deep. In mathematical terms, input is denoted as x, and output is denoted as , as shown in Figures 5–14, and can be calculated using the following equation:where represents the residual function that the block needs to learn. By rearranging equation (7), we can see that the network is learning to approximate the residual function as shown in the following equation:
The is a convolutional neural network architecture that has played a significant role in advancing the field of computer vision and image recognition. It was developed by He et al. [24]. The addresses the problem of vanishing gradients, which can often hinder the learning of intense neural networks. This problem occurs when the gradients become very small during backpropagation as they pass through layers, making it challenging for the network to learn effectively. To solve this issue, introduces residual blocks, which help the network learn identity mappings more easily, thereby allowing for the training of extremely deep networks. architecture consists of 50 layers as shown in Figure 15 and is organized in a series of building blocks, each comprising multiple convolutional layers, batch normalization, and activation functions. These blocks are repeated multiple times, with shortcut connections (also known as skip connections) that allow the gradient to flow directly through the layers, thereby mitigating the vanishing gradient problem. In summary, , Figure 15, is a groundbreaking neural network architecture that tackles the challenges of training very deep networks. By introducing residual blocks with skip connections, it enables the efficient training of networks with dozens of layers. Mathematically, employs the concept of residual functions, which helps the network learn the difference between the desired output and the input, making it easier to optimize and train deep networks effectively.
In this paper, alterations were applied to the ResNet50 architecture to adapt it for the targeted classification task, as depicted from Figure 5 to 13 , and Figure 16, as follows:(1)Exclusion of Top Classification Layer. The model, as originally designed, includes a classification layer with 1000 units, suitable for classifying images into one of the 1000 categories in the ImageNet dataset. In this paper, the top classification layer was excluded. This means that the final classification layer of the is not used.(2)Freezing Layers. After excluding the top classification layer, the code iterates through all the layers in the model and sets their trainable attribute to False. By doing so, the weights of these layers are frozen, preventing them from being updated during subsequent training. This step retains the pretrained feature extraction capabilities of while preventing them from being altered during the customization process.(3)Additional Layers for Customization. The proposed model then proceeds to add new layers to the model, which are fully connected (Dense) layers. These new layers allow you to fine-tune the model for the specific classification task. The original features are followed by these custom layers, which help capture more task-specific patterns and information.(4)Flatten Layer. A Flatten layer is inserted after the frozen layers. This layer reshapes the output from the previous layers into a one-dimensional vector, preparing the data for the subsequent fully connected layers.(5)Dense Layers. Three additional dense layers are added, each with 100 units and activation function. These layers introduce nonlinearity to the model and enable it to learn more intricate patterns and relationships in the data.(6)Output Layer. Finally, a dense layer with 2 units and , equation (6), activation function is added. This layer produces class probabilities for the two classes in a binary classification problem. The activation ensures that the predicted probabilities sum up to 1, providing a probability distribution over the classes.
Overall, the changes made to the model can be seen in Figure 16, which is a modified version of Figure 14 which involves excluding the original top classification layer, freezing the layers to preserve their pretrained features, adding custom fully connected layers for task-specific learning, and creating a new output layer for your specific classification problem. These modifications enable the utilization of the pretrained ResNet50 features, while customizing the model for the specific image classification requirements, with the architecture shown in Figure 15 and the alterations detailed in Figure 17.
ResNet50 model is specifically chosen for this problem rather than the other pretrained model because of the following:(1)Residual Connections for Gradient Flow. introduces residual connections, alleviating the vanishing gradient challenge inherent in deep neural networks. This architectural innovation facilitates smoother gradient flow during backpropagation, enabling the effective training of significantly deep models.(2)Depth Scaling and Performance Maintenance. exhibits the remarkable capability to scale in depth without encountering the degradation problem observed in shallower architectures. This phenomenon underscores the network’s resilience to diminishing performance with increasing depth.(3)Efficient Parameter Utilization. achieves competitive performance with a reduced parameter count, implying efficient extraction and utilization of hierarchical features. This attribute underscores the network’s ability to discern salient patterns with parsimonious parameterization.(4)Transfer Learning Prowess. Pretrained on expansive datasets like ImageNet, serves as a potent feature extractor, demonstrating its effectiveness in transfer learning across diverse image classification tasks. The network’s learned representations contribute significantly to enhanced generalization.(5)Architectural Adaptability. The architecture of , featuring skip connections, imparts adaptability beyond image classification. Its applicability extends seamlessly to various computer vision tasks, encompassing object detection and segmentation.(6)Benchmark Performance. consistently attains state-of-the-art performance in benchmark datasets and competitions, underscoring its efficacy in discerning intricate patterns and features within images.
Figures 5–13 and 16 provide a visual representation of the architecture of a neural network. They offer a structured and static view of how the various layers are connected and how data flow through the network. These figures illustrate the arrangement of input and output layers, hidden layers, and any branching or merging of connections. They also capture the overall structure of the model, giving a snapshot of the neural network’s design, which makes it suitable for conveying a high-level overview of the layout and connections of the model.
3.2. Suggested Optimizers
Optimization and optimizers play a crucial role in training DL models effectively [25–29]. The main objective of the optimization is to discover the collection of model parameters that can reduce a specified loss function. Optimizers refer to algorithms that aid in directing the modifications of these parameters while undergoing the training procedure.
Here is a closer look at optimization and some commonly used optimizers in DL:(1)Optimization Problem. In DL, the optimization problem [30] involves finding the values of model parameters that minimize a loss function. This loss function quantifies the difference between the model’s predictions and the actual target values. The optimization process aims to iteratively adjust the model’s parameters to reduce the loss and improve its performance on the training data.(2)Gradient Descent. Gradient descent [31–34] is the foundation of most optimization algorithms in DL. It involves calculating the gradient (partial derivatives) of the loss function with respect to each parameter and updating the parameters in the direction that reduces the loss. The general update rule is Here, the learning rate controls the step size of each update.(3)Stochastic Gradient Descent (SGD). Stochastic gradient descent [35–37] is a variation of gradient descent where instead of computing the gradient using the entire training dataset, it is computed using a randomly selected mini-batch of the data. This introduces randomness and helps the optimization process escape local minima and converge faster.(i)Optimizer Algorithms. Several optimizer algorithms [25, 26, 38, 39] enhance the basic gradient descent approach. These algorithms incorporate adaptive learning rates, momentum, and other techniques to improve convergence and training efficiency. Some popular optimizers include AdaGrad, Adam, NAdam, AdaMax, and RMSprop, which will be explained in detail in the following subsubsections.(ii)Learning Rate Scheduling. Dynamic adjustment of the learning rate during training is known as learning rate scheduling [40–43]. It involves decreasing the learning rate over time to allow the optimization process to converge more accurately. Common scheduling strategies include step decay, exponential decay, and learning rate warmup.(iii)Regularization. Regularization techniques [44–47] like L1 and L2 regularization add penalty terms to the loss function to prevent overfitting. They discourage large parameter values and promote simpler models that generalize better to new data.(iv)Second-Order Optimization. While less commonly used in DL due to computational complexity, second-order optimization methods [5, 48–50] like Newton’s method and variants use second-order information (Hessian matrix) to guide updates. These methods can converge faster but require more computational resources.
The choice of optimizer depends on the specific problem, model architecture, and dataset characteristics. Experimentation and tuning are necessary to find the most suitable optimizer and hyperparameters for a given task.
3.2.1. AdaGrad
Adaptive Gradient (AdaGrad) algorithm [51–54] is a gradient-based optimization algorithm that aims to improve the efficiency of learning in the context of training machine learning (ML) models, especially in scenarios involving sparse data and features. One of the central challenges in optimization is finding an optimal set of parameters that minimizes a given loss function. AdaGrad tackles this challenge by adjusting the learning rates for each parameter based on historical information about how the gradients have been changing over time. At its core, AdaGrad’s innovation lies in the way it adapts the learning rates for different parameters by taking into account the historical behavior of gradients. This adaptation is particularly important because traditional optimization methods often use a single learning rate for all parameters, which can lead to slow convergence or even divergence in complex optimization landscapes. Mathematically, let us break down the key steps of AdaGrad:(1)Initialization. For parameter , AdaGrad initializes a parameter vector and a variable to zero. The parameter vector represents the parameter to be optimized, and keeps track of the historical sum of squared gradients.(2)Gradient Calculation. At iteration of the optimization process, the gradient of the loss function with respect to the parameter vector is computed.
Historical Gradient Accumulation. AdaGrad’s adaptation is based on the accumulation of historical gradient information which is updated aswhere represents the gradient of the loss function with respect to parameter at iteration .
Parameter Update. The parameter is updated using the adaptive learning rate. Mathematically, is expressed aswhere represents the base learning rate, determining the step size of the update, and is a small positive constant added for numerical stability to avoid division by zero.
The critical insight of AdaGrad is that parameters with small historical gradients (indicating slow updates) will receive larger learning rate adjustments, while parameters with larger historical gradients (indicating fast updates) will experience smaller learning rate adjustments. This dynamic adjustment enables the algorithm to navigate the optimization landscape more efficiently, converging faster in regions with complex dynamics. However, while AdaGrad’s adaptivity can be advantageous, it also has its limitations. As training progresses, the historical gradient sum Gi continues to accumulate, potentially leading to overly small learning rates that hinder convergence. This phenomenon, often referred to as the “learning rate decay,” has prompted the development of variants like RMSprop and Adam, which aim to strike a balance between adaptivity and convergence stability. Overall, AdaGrad is a pioneering optimization algorithm that tailors the learning rates of parameters based on historical gradient information. By adapting learning rates to the characteristics of each parameter, AdaGrad enhances the convergence efficiency, making it particularly effective in situations where data are sparse or features have varying scales. However, its tendency to decrease learning rates over time has motivated the creation of subsequent optimization algorithms that offer improved convergence behavior while retaining the adaptivity concept.
3.2.2. Adam
Adaptive Moment Estimation (Adam) [35, 52, 55, 56] is an optimization algorithm designed to enhance the training of ML models, particularly deep neural networks. It builds upon concepts from both the momentum and Root Mean Square Propagation (RMSprop), which is described later in Section 3.2.5, optimizers to offer efficient and adaptive updates to model parameters during the training process. Adam’s innovation lies in its ability to adjust both the learning rates and momentum for each parameter based on historical gradient information, leading to improved convergence properties in various optimization landscapes. Mathematically, let us delve into the fundamental aspects of the Adam optimizer:(1)Initialization. Adam initializes two moment vectors, and , for each parameter. The vector tracks the exponentially moving average of past gradients, and the vector tracks the exponentially moving average of squared past gradients.(2)Exponential Moving Averages. In each iteration of training, the gradients of the loss function concerning the parameters are calculated. Adam then updates the and vectors using equations (12) and (13). where and are hyperparameters that control the exponential decay rates of the moving averages.(3)Bias Correction. The and vectors are biased toward zero, especially during the initial iterations. To address this bias, at iteration , Adam performs bias correction by adjusting the vectors using equations (14) and (15):(4)Parameter Update. In the final step, the model’s parameters is updated utilizing the bias-corrected moment estimates as described in equation (16):where is the learning rate, controlling the step size of the update, and is a small constant added for numerical stability to prevent division by zero.
Adam’s key strength lies in its adaptive learning rates and momentum adjustments. The vector behaves like momentum, helping to accelerate the optimization process, while the vector adapts the learning rates based on the historical information of the squared gradients. This combination allows Adam to handle sparse gradients and varying landscape geometries effectively. However, it is worth noting that while Adam offers excellent performance in many scenarios, its hyperparameters (, , and ) need careful tuning. Incorrect tuning can result in suboptimal convergence or training instability. Furthermore, in cases where the loss landscape is particularly noisy or ill-conditioned, variations of Adam or other optimizers might be more suitable. Overall, Adam is an advanced optimization algorithm that combines momentum and adaptive learning rates to optimize ML models effectively. Its mathematical framework enables it to adaptively adjust learning rates and momentum, making it well-suited for training deep neural networks and handling challenges like sparse gradients and varying scales. However, its success relies on hyperparameter tuning and understanding the characteristics of the optimization landscape.
3.2.3. AdaMax
Adaptive Moment Estimation with Infinity Norm (AdaMax) [34, 57–59] is an optimization algorithm that builds upon the principles of the Adam optimizer while addressing some of its limitations. Specifically, AdaMax is designed to offer more stable and effective updates by modifying the way the moments (exponential moving averages of gradients and squared gradients) are computed. By introducing a new term called “infinity norm” and modifying the denominator of the update equation, AdaMax aims to mitigate some of the potential issues that can arise during optimization. Let us explore the mathematical operations and key concepts of AdaMax:(1)Initialization. Just like other adaptive optimization algorithms, AdaMax initializes two moment vectors, and , for each parameter. These moments track the exponentially weighted averages of past gradients and squared gradients, respectively. Additionally, an initial value for the infinity norm is set.(2)Exponential Weighted Averages. At iteration , the gradients of the loss function with respect to the parameters are calculated. AdaMax then updates the moments and using equations (17) and (18): where and are the exponential decay rates for the moving averages.(3)Parameter Update. AdaMax performs the parameter update using the modified moments:where is the learning rate, controlling the step size of the update.
The key innovation of AdaMax lies in the computation of the moment. Instead of using the squared gradient directly, AdaMax employs the infinity norm , which is the maximum absolute value of the gradient components. By doing so, AdaMax introduces a form of regularization on the second moment that can help address some of the convergence issues observed in Adam when the moving average of squared gradients gets close to zero. Moreover, AdaMax’s use of the infinity can make the updates more robust, particularly in cases where individual gradient components are exceptionally large. This adaptation can prevent the moment from becoming too small, thus maintaining stable learning rates throughout training. Overall, AdaMax is an extension of the Adam optimizer that employs the infinity norm and a modified denominator in the update equation to provide more reliable and effective parameter updates. Its mathematical alterations are designed to address some of the potential challenges in optimization landscapes, making it a valuable optimization technique for training ML models, especially deep neural networks. However, like other adaptive optimizers, proper hyperparameter tuning is essential to harness its full potential.
3.2.4. NAdam
Nesterov-accelerated Adaptive Moment Estimation (NAdam) [34, 60–62] is an optimization algorithm that builds upon the Adam optimizer by incorporating Nesterov’s accelerated gradient technique. NAdam aims to combine the benefits of both the Nesterov Accelerated Gradient (NAG) and Adam optimizers to improve convergence and handling of noisy or ill-conditioned optimization landscapes. By adopting the momentum update in NAG with the moment-based update in Adam, NAdam seeks to achieve enhanced optimization performance. To understand NAdam more thoroughly, let us delve into its mathematical operations and key concepts:(1)Initialization. Like other adaptive optimization algorithms, NAdam initializes two moment vectors, and , for each parameter. These moments track the exponentially weighted averages of past gradients and squared gradients, similar to Adam. Additionally, NAdam also initializes the Nesterov momentum term for each parameter.(2)Exponential Weighted Averages. In each iteration of training, the gradients of the loss function with respect to the parameters are calculated. NAdam then updates the moments m and using similar formulas as Adam in equations (20) and (21).(3)Nesterov Momentum Update. NAdam introduces the Nesterov momentum term into the momentum update step. This term incorporates information from the previous iteration’s gradient as expressed in the following equation:(4)Combined Parameter Update. The NAdam parameter update combines the momentum update from Nesterov’s technique and the moment-based update from Adam. The parameter is updated using the following equation:
The strength of NAdam lies in its combination of Nesterov’s accelerated gradient and the adaptive moment updates. By incorporating the Nesterov term into the momentum update, NAdam improves convergence properties by adjusting the direction of the update based on past gradients. This can be particularly beneficial when navigating through regions with complex and noisy landscapes. Overall, NAdam is an optimization algorithm that marries the concepts of Nesterov’s accelerated gradient and the adaptive moment-based updates of Adam. Its hybrid approach aims to provide improved optimization performance, making it a suitable candidate for training ML models, especially deep neural networks. As with other adaptive optimizers, proper tuning of hyperparameters is crucial for achieving optimal results.
3.2.5. RMSprop
Root Mean Square Propagation (RMSprop) [23, 34, 63–68] is an optimization algorithm designed to address some of the limitations of traditional gradient descent methods, particularly in handling varying scales of gradients and accelerating convergence in DL models. RMSprop adapts the learning rates of individual parameters based on the historical information of past gradients, allowing for more efficient updates and better optimization performance. Let us explore the mathematical operations and key concepts of RMSprop:(1)Initialization. RMSprop initializes a variable s for each parameter, representing the exponentially moving average of squared gradients. This variable is initialized to zero.(2)Exponential Moving Averages. In each iteration of training, the gradients of the loss function with respect to the parameters are calculated. RMSprop then updates the s values using the following equation: where β is a hyperparameter that controls the exponential decay rate of the moving average.(3)Parameter Update. RMSprop performs the parameter update using the adapted learning rate.
The main innovation of RMSprop lies in its adaptation of the learning rate based on the moving average of squared gradients. By doing so, RMSprop can effectively handle situations where some gradients have larger magnitudes than others. This adaptive learning rate adjustment allows the optimizer to navigate optimization landscapes more smoothly and converge faster. Moreover, RMSprop’s incorporation of the squared gradient magnitude in the denominator of the update equation contributes to a natural scaling of the learning rates, preventing the learning rates from becoming too small, as can happen in standard gradient descent methods. This property can make RMSprop particularly useful for training deep neural networks, where gradients can exhibit a wide range of magnitudes. Overall, RMSprop is an optimization algorithm that adapts the learning rates of parameters based on the historical information of squared gradients. Its mathematical framework enables it to handle varying scales of gradients and accelerate convergence in DL models. While RMSprop has been proven effective in many scenarios, it is important to note that proper hyperparameter tuning is crucial for achieving optimal results, as the choice of β and ϵ can significantly impact the optimization process.
4. Analytical Insights and Experimental Outcomes
The proposed models along with the recommended optimizers were subjected to experimentation, the datasets used for both training and testing were evaluated, and the final results were obtained by averaging all the assessment metrics. Additionally, the time taken for training and testing was also recorded. Further details on the datasets used to evaluate the performance of the proposed model can be found in Section 4.1. Workplace characteristics and hyperparameter values are recorded in Section 4.2, followed by performance indicators in Sections 4.3 and 4.4, which describe experimental analysis and result discussion.
4.1. Dataset Description
The evaluation of the proposed and state-of-the-art models is conducted in this study using data from a CXR provided by the NIH clinical center. This CXR can give an initial indication of cardiomegaly by evaluating the size and shape of the heart. NIH, a research hospital in the United States, has recently published over 100,000 anonymized frontal-view CXR images. The NIH gathered a scanning dataset of over 30,805 individuals, many of whom had advanced lung illnesses. All of the photographs have a high resolution of 1024 × 1024. Figure 18 provides visual dataset samples. There are up to 15 different thoracic categories, such as atelectasis, cardiomegaly, effusion, infiltration, mass, nodule, pneumonia, pneumothorax, consolidation, edema, emphysema, fibrosis, pleural thickening, hernia, and normal images. Although the picture label is derived using natural language processing (NLP), the publisher guarantees that the NLP labeling precision surpasses 90%. The CXR8 also provides 8 class designations. The dataset is offered by Kaggle but with some changes to be more suitable for our problem of early detection of cardiomegaly and to make the data more balanced. The dataset is composed of 5552 images, divided into two categories: 2776 images are labeled as No Cardiomegaly, indicating normal conditions, and the other 2776 images are categorized under Cardiomegaly, signifying the presence of the disease, as illustrated in Figure 19. These images are preprocessed through resizing, color transformation, and normalization and then labeled. Image preprocessing is a crucial initial step in computer vision and image analysis pipelines, aiming to enhance the quality and consistency of images prior to any further analysis or processing.
4.1.1. Resizing [69–71]
The first step is to adjust the image’s dimensions while keeping its aspect ratio intact. This is necessary to make all the images the same size, which ensures consistency and compatibility for further processing. Resizing images not only makes computing easier but also helps to compare images fairly across various resolutions, making analysis more accurate.
4.1.2. Color Transformation [72, 73]
In the second step, the color space of an image is manipulated to enhance certain image characteristics. For example, RGB (Red, Green, Blue) and HSV (Hue, Saturation, Value) are distinct color spaces that represent color information within an image. Transforming an image from one color space to another can accentuate or suppress certain features, which can be useful in tasks such as feature extraction or pattern recognition.
4.1.3. Normalization [74, 75]
The third step involves adjusting pixel values to ensure they comply with a standardized scale. This is important to reduce the impact that changing lighting conditions or sensor characteristics can have on the appearance of an image. Normalization techniques typically involve scaling pixel values to fall within a specific range, such as [0, 1] or [−1, 1], or performing mean subtraction and division by standard deviation. Normalization ensures that the statistical distribution of image data is consistent, which helps to stabilize the training and learning processes of subsequent ML models.
4.1.4. Labeling [76, 77]
The final step in the process involves annotating images with relevant information that identifies the objects or regions within them. In supervised learning, labels serve as the ground truth for training and evaluating ML models. Proper labeling requires accurately associating images with the appropriate classes or categories, allowing algorithms to effectively learn and generalize from the provided data.
In this study, the data are labeled as 0 for no cardiomegaly and 1 for cardiomegaly. The data are then shuffled and split into 80% training and 20% testing sets, as shown in Table 2.
Next, the dataset is enhanced with data augmentation using the ImageDataGenerator technique [78–83]. Data augmentation is a crucial technique in computer vision, especially when there is a shortage of annotated data for training ML models. This technique generates various versions of the existing dataset by applying a range of transformations to the original images. In the Keras library, the ImageDataGenerator class enables this process by providing a simple way to implement data augmentation during the model’s training phase.
The ImageDataGenerator technique [80–83] operates by applying a set of predefined image transformations, such as rotations, flips, shifts, zooms, and shears, to the input images. These transformations introduce controlled perturbations that simulate real-world variations in imaging conditions. By augmenting the dataset with these transformed images, the effective size of the training data increases substantially, which can mitigate overfitting and enhance the model’s generalization capabilities.
The advantages of data augmentation [84] are manifold. Firstly, augmented data enrich the training dataset with diverse instances, effectively expanding the model’s exposure to a wider range of scenarios. This exposure contributes to improved robustness, as the model learns to recognize objects or patterns despite variations in lighting, orientation, and other factors. Secondly, augmented data can mitigate the risk of overfitting, a phenomenon where a model learns to memorize the learning data instead of generalizing from it. The inherent noise introduced by the transformations helps regularize the learning process, preventing the model from becoming excessively specialized to the training samples. Lastly, data augmentation can also save valuable time and resources, as it reduces the dependency on collecting and annotating an extensive dataset. Through augmentation, a relatively smaller dataset can be effectively amplified, potentially leading to comparable or even superior model performance.
4.2. Modelling Context
The computational outcomes were produced using a computer equipped with an Intel Core i7 CPU, 64 GB of RAM, and an NVIDIA GTX 1050i GPU. Additionally, coding resources like Python, Keras, TensorFlow, and Sklearn were employed for executing the programming assignments. The hyperparameters and standard parameter configurations for the presented models, including the maximum number of epochs and the loss function, are detailed in Table 3. These hyperparameters were set by using both grid search and random search techniques for hyperparameter tuning. Grid search exhaustively explores a predefined set of hyperparameter values, as the number of epochs was in the range from 1 to 100, the learning rate was in the range from 0.0001 to 0.1, and batch size was in the range from 16 to 120, while random search randomly samples from a given distribution. This combination helps find optimal configurations more efficiently.
4.3. Performance Metrics
Recall, precision, accuracy, and F1-score are some of the assessment measures used to assess the presented models. These metrics are dependent on the assessment parameters for predictive models known as True Positive (), True Negative (), False Positive (), and False Negative (). Recall [85] is mathematically defined as divided by the product of and , as shown in the following equation:
Precision [86] is computed using equation (25) and is equal to divided by the product of and :
Equation (26) describes accuracy as follows:
F1-score [87] is equal to 2 times precision times recall times the sum of precision and recall divided by the precision plus recall. It is written as
4.4. Investigative Comparison
This paper suggests a framework to address the problem of early cardiomegaly detection. Two different DL models were used for this objective and each one is developed with five types of optimizers, which are AdaGrad, Adam, AdaMax, NAdam, and RMSprop, to find the best optimizer for each proposed model. The goal is to develop two DL models and find the best optimizer from the fifth for each one of them that performs at its best regarding processing speed and detection accuracy and then find the best-proposed model from the two.
4.5. Results of the Proposed CNN Models
In this section, the simulated outcomes of the suggested CNNs with the five types of optimizers are compared to find the best optimizer for the proposed CNN model. Table 3 shows that the suggested CNNs with the five types of optimizers are all trained using 50 epochs with a batch size of 85.
4.5.1. Results of the Proposed CNN with AdaGrad
The simulated outcomes of the proposed CNN-AdaGrad are covered in this section. The accuracy, loss curves, and confusion matrix of the proposed CNN-AdaGrad are shown in Figures 20–22, respectively. The proposed CNN-AdaGrad obtained an accuracy of 96.07% for training. Table 4 shows the proposed CNN-AdaGrad evaluation metrics in the testing stage. It achieved an accuracy, precision, recall, and F1-score of 95.5%, 96.89%, 73.1%, and 83.3%, respectively.
4.5.2. Results of the Proposed CNN with Adam
This section covers the simulated outcomes of the proposed CNN-Adam. The accuracy, loss curves, and confusion matrix of the proposed CNN-Adam are shown in Figures 23–25, respectively. The proposed CNN-Adam obtained an accuracy of 100.00% for training. Table 5 shows the proposed CNN-Adam evaluation metrics in the testing stage. It achieved an accuracy, precision, recall, and F1-score of 99.64%, 100.00%, 97.39%, and 98.7%, respectively.
4.5.3. Results of the Proposed CNN with AdaMax
This section covers the simulated outcomes of the proposed CNN-AdaMax. The accuracy, loss curves, and confusion matrix of the proposed CNN-AdaMax are shown in Figures 26–28, respectively. The proposed CNN-AdaMax obtained an accuracy of 100.00% for training. Table 6 shows the proposed CNN-AdaMax evaluation metrics in the testing stage. It achieved an accuracy, precision, recall, and F1-score of 99.91%, 100.00%, 99.40%, and 99.7%, respectively.
4.5.4. Results of the Proposed CNN with NAdam
The simulated outcomes of the proposed CNN-NAdam are covered in this section. The accuracy, loss curves, and confusion matrix of the proposed CNN-NAdam are shown in Figures 29–31, respectively. The proposed CNN-NAdam obtained an accuracy of 100.00% for training. Table 7 shows the proposed CNN-NAdam evaluation metrics in the testing stage. It achieved an accuracy, precision, recall, and F1-score of 99.55%, 100.00%, 97.1%, and 98.5%, respectively.
4.5.5. Results of the Proposed CNN with RMSprop
The simulated outcomes of the proposed CNN-RMSprop are covered in this section. The accuracy, loss curves, and confusion matrix of the proposed CNN-RMSprop are shown in Figures 32–34, respectively. The proposed CNN-RMSprop obtained an accuracy of 99.50% for training. Table 8 shows the proposed CNN-RMSprop evaluation metrics in the testing stage. It achieved an accuracy, precision, recall, and F1-score of 99.37%, 98.9%, 96.89%, and 97.8%, respectively.
4.6. Results of the Modified ResNet50 Models
In this section, the simulated outcomes of the modified ResNet50 with the five types of optimizers are compared to find the best optimizer for the modified ResNet50 model. Table 3 shows that the modified ResNet50 models with the five types of optimizers are all trained using 15 epochs with a batch size of 32.
4.6.1. Results of the Modified ResNet50 with AdaGrad
The simulated outcomes of the modified ResNet50-AdaGrad optimizer are presented here. The accuracy, loss curves, and confusion matrix of the modified ResNet50-AdaGrad are shown in Figures 35–37, respectively. The modified ResNet50-AdaGrad obtained an accuracy of 100.00% for training. Table 9 shows the modified ResNet50-AdaGrad evaluation metrics in the testing stage. It achieved an accuracy, precision, recall, and F1-score of 99.73%, 99.4%, 98.9%, and 99.1%, respectively.
4.6.2. Results of the Modified ResNet50 with Adam
The simulated outcomes of the modified ResNet50-Adam are covered in this section. The accuracy, loss curves, and confusion matrix of the modified ResNet50-Adam are shown in Figures 38–40, respectively. The modified ResNet50-Adam obtained an accuracy of 100.00% for training. Table 10 shows the modified ResNet50-Adam evaluation metrics in the testing stage. It achieved an accuracy, precision, recall, and F1-score of 99.19%, 100.00%, 95.1%, and 97.5%, respectively.
4.6.3. Results of Modified ResNet50 with AdaMax
The simulated outcomes of the modified ResNet50-AdaMax are covered in this section. The accuracy, loss curves, and confusion matrix of the modified ResNet50-AdaMax are shown in Figures 41–43, respectively. The modified ResNet50-AdaMax obtained an accuracy of 100.00% for training. Table 11 shows the modified ResNet50-AdaMax evaluation metrics in the testing stage. It achieved an accuracy, precision, recall, and F1-score of 99.64%, 100.00%, 97.50%, and 98.70%, respectively.
4.6.4. Results of Modified ResNet50 with NAdam
The simulated outcomes of the modified ResNet50-NAdam optimizer are covered in this section. The accuracy, loss curves, and confusion matrix of the modified ResNet50-NAdam are shown in Figures 44–46, respectively. The modified ResNet50-NAdam obtained an accuracy of 100.00% for training. Table 12 shows the modified ResNet50-NAdam evaluation metrics in the testing stage. It achieved an accuracy, precision, recall, and F1-score of 99.46%, 100.00%, 96.6%, and 98.3%, respectively.
4.6.5. Results of Modified ResNet50 with RMSprop
The simulated outcomes of the modified ResNet50-RMSprop optimizer are covered in this section. The accuracy, loss curves, and confusion matrix of the modified ResNet50-RMSprop are shown in Figures 47–49, respectively. The modified ResNet50-RMSprop obtained an accuracy of 98.96% for training. Table 13 shows the modified ResNet50-RMSprop evaluation metrics in the testing stage. It achieved an accuracy, precision, recall, and F1-score of 96.94%, 99.3%, 80.2%, and 88.7%, respectively.
4.6.6. Result Discussion
Table 14 shows the experimental results of the proposed CNN with the suggested optimizers; it can be observed that AdaGrad optimizer is the last in all performance measures. Adam optimizer ranks first in precision and second in accuracy, recall, F1-score, and loss. AdaMax optimizer is the best in all performance measures, and it is also the lowest in loss. NAdam optimizer ranks first in precision and third in accuracy, recall, F1-score, and loss. RMSprop optimizer ranks second in precision and fourth in accuracy, recall, F1-score, and loss.
Based on the duration required for the proposed CNN model, it is apparent that during the training phase, AdaGrad, Adam, and RMSprop optimizers required equal time and ranked second. AdaMax and NAdam, on the other hand, took the least amount of time and thus ranked first. Moving on to the testing phase, AdaMax proved to be the most efficient, taking the shortest time. NAdam came in second, while Adam and RMSprop took the same amount of time and ranked third. Finally, AdaGrad consumed the most time and ranked last.
Table 15 shows the experimental results of the modified ResNet50 with the suggested optimizers. It can be observed that AdaGrad optimizer ranks first in accuracy, recall, and F1-score and second in precision and loss. Adam optimizer ranks first in precision and fourth in accuracy, recall, F1-score, and loss. AdaMax optimizer ranks first in precision and loss and second in accuracy, recall, and F1-score. NAdam optimizer ranks first in precision and third in accuracy, recall, F1-score, and loss. RMSprop is the last in all performance measures.
It is evident from the duration needed for the adjusted ResNet50 model that Adam and RMSprop optimizers took an equal amount of time and were second in rank during the training phase. In contrast, AdaGrad, AdaMax, and NAdam took the least amount of time and ranked first. During the testing phase, AdaGrad demonstrated the highest efficiency, taking the shortest time. NAdam followed in second place, while AdaMax came in third. Adam and RMSprop took the same amount of time and ranked last.
Finally, it can be observed that the proposed CNN-AdaMax is the best in all performance measures, faster in both training and testing time, and lowest in loss.
4.7. Benchmarking against Cutting-Edge Models
Table 16 compares the proposed models with the recent models using the same dataset. The proposed CNN-AdaMax optimizer ranked first in all performance metrics, while the proposed CNN-Adam optimizer, the proposed CNN-NAdam optimizer, the modified ResNet50-Adam optimizer, the modified ResNet50-AdaMax optimizer, and the modified ResNet50-NAdam optimizer ranked first in precision. The modified ResNet50-AdaGrad ranked second in terms of accuracy, precision, recall, and F1-score.
It is important to note that the choice of optimizer depends on various factors, including the architecture of the neural network, the nature of the dataset, and the training goals. In this study, the AdaMax optimizer with the proposed CNN, utilized dataset, and same hyperparameters is the best in all performance measures because as previously explained in Section 3.2.3, AdaMax is a variant of the Adam optimizer that provides certain advantages, particularly for training the proposed CNN:(1)Sparse Gradient Handling. AdaMax’s parameter update rule includes the infinity norm (L∞ norm) of the gradients, which can be helpful for handling sparse gradients. This property can be beneficial when training deep models with a large number of parameters, as sparse gradients can be an issue during optimization.(2)Convergence Speed. AdaMax is known to have better convergence properties compared to plain Adam in certain cases. This can be beneficial when training a CNN from scratch, as faster convergence can lead to quicker training times.(3)Stability. AdaMax’s formulation can make it more stable than other optimizers, such as plain Adam. This stability can help avoid oscillations during training and improve the overall training process.
On the other hand, AdaGrad with the modified ResNet50 model in our problem case and with the same hyperparameters is the first in terms of accuracy, recall, and F1-score and also is faster in training and testing time because as previously explained in Section 3.2.1, AdaGrad is an optimizer that adapts the learning rates of individual parameters based on their historical gradient information. It has certain characteristics that could be advantageous when training complex architectures like ResNet50:(1)Adaptivity to Sparse Features. AdaGrad’s adaptive learning rate mechanism works well with sparse features. ResNet50 has skip connections that can introduce sparse-like activations through the network. AdaGrad’s adaptivity to such features can lead to more effective updates during training.(2)Weight Decay Adaptation. AdaGrad’s accumulation of historical gradients effectively reduces the learning rate for frequently updated parameters. This can act as a form of implicit weight decay, which can help prevent overfitting in complex models like ResNet50.(3)Convergence with Sparse Data. If the dataset used for training ResNet50 contains sparse or irregularly distributed features, AdaGrad’s learning rate adaptation can aid in navigating the optimization landscape and achieving convergence.
Overall, it is important to note that while AdaMax and AdaGrad or other optimizers have these potential advantages, their effectiveness can still depend on factors such as the specific dataset, hyperparameter tuning, and other architecture-related choices. So, in order to find out which optimizer is most effective for a particular task, experimentation is crucial.
5. Conclusion
Early detection of cardiomegaly is crucial for managing cardiovascular health. This research highlights the importance of timely identifying the condition to enable proactive interventions, ultimately leading to better patient outcomes. Advanced imaging techniques, such as CXR imaging, are noninvasive and effective for assessing cardiac morphology and size. They enable clinicians to determine cardiac chamber enlargement and overall cardiac dimensions accurately. So, in our paper, one of the finest datasets for the detection process is CXR. DL-based models are used to categorize images in the medical area. This paper introduced two DL models: one from scratch which is the proposed CNN and a pretrained model ResNet50 with some modifications, to be suitable for our binary classification problem. Each model is compiled with five types of optimizers which are AdaGrad, Adam, AdaMax, NAdam, and RMSprop. For the proposed CNN which is trained using 50 epochs with a batch size of 85, AdaMax optimizer is the best in all performance measures and it is also the fastest of them. For modified ResNet50 which is trained using 15 epochs with a batch size of 32, the AdaGrad optimizer is better in terms of accuracy, recall, and F1-score and also is the fastest of them, but Adam, AdaMax, and NAdam optimizers are the better in precision. Overall, the proposed CNN-AdaMax is the best in all performance measures and the fastest. The results of this study confirm the diagnostic power of such imaging modalities in detecting cardiomegaly. Therefore, they should be considered for integration into routine clinical practice for early detection and intervention. Early detection of cardiomegaly not only advances scientific understanding but also reduces the burden of advanced cardiac pathology on individuals and healthcare systems. Further research should investigate the long-term implications of early detection and explore the interplay of various imaging technologies. This will contribute to the refinement of clinical strategies and the enhancement of patient care in cardiovascular health. Although the study achieved impressive accuracies of 99.91% using AdaMax for the proposed CNN and 99.73% using AdaGrad for the modified ResNet50, there are still some limitations. The dataset used for this study includes 5552 images, with equal representation of no-cardio and cardio cases, but there may be biases that could affect the generalizability of the results. Additionally, the complex architecture of ResNet50 makes it difficult to interpret the model’s results, which is crucial in medical applications. The high resource requirements for training ResNet50 also raise concerns about practical deployment. To overcome these limitations, future research will focus on exploring the dataset in detail to identify and address any biases, working collaboratively with domain experts to improve interpretability, and investigating optimization strategies such as model compression and hardware acceleration to enable efficient deployment in resource-constrained environments.
Abbreviations
| DL: | Deep learning |
| ResNet50: | Residual network with 50 layers |
| NIH: | National Institutes of Health |
| AdaGrad: | Adaptive Gradient |
| NAdam: | Nesterov-accelerated Adaptive Moment Estimation |
| RMSprop: | Root Mean Square Propagation |
| AdaMax: | Adaptive Moment Estimation with Infinity Norm |
| VGG19: | Visual Geometry Group with 19 layers |
| AI: | Artificial intelligence |
| CTR: | Cardiothoracic ratio |
| CXRDANet: | Chest X-ray with dual attention network |
| PA: | Posterior-anterior |
| AUC: | Area under the curve |
| GAP: | Global average pooling |
| SGD: | Stochastic gradient descent |
| CNN-Adam: | CNN with Adam optimizer |
| CNN-NAdam: | CNN with NAdam optimizer |
| ResNet50-AdaGrad: | ResNet50 with AdaGrad optimizer |
| ResNet50-AdaMax: | ResNet50 with AdaMax optimizer |
| ResNet50-RMSprop: | ResNet50 with RMSprop optimizer |
| ML: | Machine learning |
| CNNs: | Convolutional neural networks |
| VGG16: | Visual Geometry Group with 16 layers |
| Adam: | Adaptive Moment Estimation |
| TN: | True Negative |
| FP: | False Positive |
| FN: | False Negative |
| TP: | True Positive |
| NLP: | Natural language processing |
| CXR: | Chest X-ray |
| CAM: | Channel attention module |
| SAM: | Spatial attention module |
| AMM: | Attention mapping mechanism |
| JSRT: | Japanese Society of Radiological Technology |
| CNN-AdaGrad: | CNN with AdaGrad optimizer |
| CNN-AdaMax: | CNN with AdaMax optimizer |
| CNN-RMSprop: | CNN with RMSprop optimizer |
| ResNet50-Adam: | ResNet50 with Adam optimizer |
| ResNet50-NAdam: | ResNet50 with NAdam optimizer. |
Data Availability
Data are available on request from the corresponding author.
Disclosure
Intellectual property protection measures have been thoroughly discussed, and there are no impediments to its dissemination, including publication timing. Accordingly, we confirm our adherence to our institutions’ guidelines on intellectual property.
Conflicts of Interest
The authors declare that they have no conflicts of interest.
Authors’ Contributions
All authors have reviewed and approved the manuscript, and no supplementary contributors who meet authorship criteria have been excluded. The sequence of authors has been agreed upon mutually.
Acknowledgments
The authors extend their sincere thanks to the Deanship of Scientific Research, Vice Presidency for Graduate Studies and Scientific Research, King Faisal University, Saudi Arabia, for their generous financial support, which was crucial for the research project (GRANT 5,740).