|
Segmentation techniques | Overview | Advantages | Drawbacks |
|
Thresholding [4] | This method is based on the threshold maximum and the minimum value, corresponding to different peaks depicting different regions. Various techniques have emerged from high threshold values like a balanced histogram, k-means, and otsu maximum variance | No prerequisites are required about image and computation is fast with less complexity | Information with low peaks is not considered and ignored; hence, continuous value is not obtained. In the presence of noise and poor contrast, performance is not up to mark |
|
Region-based [16, 27] | Identical regions are grouped using techniques like region growing, splitting, merging, etc | Better than edge detection about noise immunity. It works better in homogeneous regions | Quite expensive in the context of both time and memory |
|
Clustering [10] | It creates different clusters in the spatial domain. Groups are homogeneous | Best results on overlapped data. Results best for classification. Suitable for real-life applications as it uses fuzzy logic | Clustering validity is a challenging task to be determined. Expensive and sensitive to primary clusters |
|
Edge detection [22] | Work on discontinuity principle and locate regions with minimum sudden change. It is of two types sequential and parallel | Results are better with images having a high contrast value | Give unexpected results with images having many edges or improperly defined boundaries and noisy images |
|
Contour-based segmentation [25] | In this work computer-aided diagnosis (CAD) and fusion via CNN is done for recognition, analysis, and further treatment with the help of RF giving maximum accuracy of 97.51% and minimum error via CNN classifier 95.65%. After segmentation using sparse transform, the algorithm eliminates the attributes of the point spread function | Diagnosis of cancer and accuracy detection is now upgraded with machine learning techniques. This work offers enhanced performance and better implementation results, with accuracy and lesser time in medical CT image restoration, new image recovery with quality and speed | The present CAD system for measuring the accuracy is not recognized and acceptable |
|
Energy function-based technique [28] | It is based on parameterizing the curve, taking some sampling values | Minimum processing is required and if flexible. Fast and efficient | It is not useful in the case of higher dimensions, selection of sampling strategy, topology changes, etc |
|
Breast-region segmentation [29] | This segmentation has three divisions: Classical segmentation, which includes region, threshold, edge-based segmentation utilizing supervised and unsupervised methods using deep learning | U-net was chosen because, unlike other deep learning models, it does not require annotated photos | The method is easily simulated by AI, ML, or DNN persons rather than physicians or biologists |
|
Optimized-region growing technique [30] | The optimal features are selected by a hybrid optimization algorithm and classified using a neural network | The velocity updated lion algorithm combines the lion algorithm (LA) with PSO to achieve the best feature selection and weight optimization using NN (VU-LA) and VU-LA is also compared to existing models such as the whale optimization algorithm, grey wolf optimization, firefly, PSO, and LA in terms of performance | The algorithm cannot work well on hazy or faint images |
|
Region of interest [31] | After segmentation to obtain a region of interest (ROI), the images were cleaned up using a median filter and compared using ANN, SVM, and reduced features of SVM | The statistics and grey level cooccurrence matrix are utilized to classify to extract the features from enhanced images using the hybrid SVM-ANN | A better algorithm can be made for enhanced feature detection |
|
Supervised segmentation [32] | The work proposes to constrain the segmentation output when morphological operations to measure performance which uses top-hat and closing operations to evaluate on high-resolution images from the INBreast dataset | It achieves an increase in F1 and in the recall if compared to the training without morphology loss | The evaluations cannot be justified sometimes when images are taken from different sources |
|
Pseudo-color segmentation [33] | Thermal cameras record radiation images, which are then transformed to images of pseudo-colored. All the colors of the thermogram correspond to a specific temperature. The interpretation of breast thermograms is mostly dependent on color analysis and thermogram asymmetry analysis. The work analyses breast thermograms by segmenting the RoI, extracted as a hot region, and then analyzing the color. Abnormalities are shown in the hottest regions by contours | The results compared to diagnosis to ensure infrared thermography is a reliable tool for detecting breast cancer | Sometimes pathologists fail to identify the raw image received by the radiation |
|
Graph segmentation [34] | The automatic segmentation of stained tissue images aid in the detection of malignant disease, and this is done after the separation of contacting cells. Traditional segmentation algorithms face numerous challenges. They present a novel automatic approach for segmenting clustered cancer cells in this work. In the first stage, also use Chan-Vese energy functional to determine cell areas by a modified geometric active contour | By determining high concavity locations along with the cell outlines, contacting cell areas are recovered from the presegmented image | Image profiles can sometimes fail to subclassify breast tumours into additional subtypes, which can help in diagnosis and survival |
|
Variant feature transformation [35] | To subclassify extremely aggressive breast malignancies, the researchers used public transcriptomics datasets in breast cancer cell lines and breast cancer tumours, and associated splice variants | Splicing is becoming more often used as a biomarker for grading tumours | Sometimes splicing may not be an exact method to classify their variants |
|
Region –growing algorithm [36] | The work provides a new CRG (conditional region growth) approach for determining correct MC bounds beginning from seed points selected, and they are determined by detecting regional maxima and analyzing superpixels. The region-growing stage is then maintained by the criteria set tuned to MC detection for contrast and shape variation obtained from prior knowledge, which determines the size of the searching area in the neighbourhood. To do qualitative and quantitative analysis for detection of MC and delineation many experiments are done on MC of multiple types | The importance of utilized criteria in the context of MC delineation for better management of breast cancer is demonstrated by a comparison of the proposed technique state-of-the-art | Microcalcifications with their morphology are the indicators of breast cancer when shape and size are considered when malignity degree is to be found out and therefore delineation of MC is done for diagnosis of cancer |
|
Graph-cut algorithm [37] | The mammograms can be classified with BIRADS which subjects to qualitative assessments and face inter-reader and intra-reader variations. Comparative analysis of various articles is done for classification accuracy and computational complexity to design an algorithm for the measurement of breast density using machine learning or deep learning | Breast density measurement via machine learning and deep learning is increasing the rapid development | If the density of the breast increases, then chances of breast cancer also increase which reduces the sensitivity of measuring mammographic density |
|
Watershed algorithm [38] | A novel approach for detection of breast cancer automatically from histopathological images that are composed of binary and eight-class based on a convolutional-LSTM learning model trained on BreakHis dataset and preprocessing using marker-controlled watershed segmentation algorithm and optimized SVM classifier. When MWSA is used with the optimized SVM classifier with Bayesian optimization processed HPIs improve when compared to the CLSTM model’s softmax classifier | When compared to present approaches utilizing BreakHis dataset, the methodology achieves great performance for both classifications | Sometimes histopathological images cannot be classified and may not detect breast cancer, therefore, AI and deep learning-based applications are used for automated breast cancer detection with high performances |
|
Fuzzy –C means [39] | This method uses computer-aided cancer diagnosis in its various stages. The first stage performs a reduction of noise and preprocessing contrast enhancement, while the 2nd stage segments the ROI and, the kernel fuzzy C-means method is used for segmentation, and features are extracted and feature selection is employed. For final identification, these selected features are fed into an SVM | The proposed version of the metaheuristic neural network optimization approach improves feature selection and SVM classifier performance. When the methodology was compared against five state-of-the-art methodologies, results revealed the approach was superior | Sometimes Computer-aided design lacks when input is not appropriate |
|
Otsu’s optimal thresholding [40] | The paper creates a framework that scans the stage of cancer by using the optimized kernel fuzzy clustering algorithm to determine cancer and identify segmented regions in mammogram images. The mammographic images are preprocessed noise-free images obtained by using the hybrid denoising filter algorithm. Data clustering is done to classify data of similar types in one group and of dissimilar types in another group | Results give accuracy & efficiency of the proposed system compared to methods such as K-means, OKFCA, and otsu | Issues like mammograms deviate artifacts, are similar breast tissues and contrast issues on the boundary between skin and air |
|
Fusion of K-means and region growing algorithms [41] | An automated system used by radiologists for diagnostic decisions involving detection of breast masses by using an optimized region growing method, optimal seed point selection, and threshold generation were achieved using grey wolf optimization. The features for both global and local are extracted for shape features, grey level: co-occurrence matrix, run length matrix, texture feature: Local binary pattern, and scale-invariant feature transform | An amalgamation of local and global features with an SVM classifier differentiates benevolent or malignant images and an accuracy of 96% is achieved by GLCM and LBP | The existence rate after detection of cancer affected person cannot be sometimes predicted |
|