Artificial Neural Networks in Mammography Interpretation and Diagnostic Decision Making

Table 1

Summary of ANN studies in mammography interpretation and diagnostic decision making.

Study	Type	ANN structure	Input	Dataset and training/testing strategy	Results and findings
Stafford et al. (1993) [33]	CADe	A committee of four three-layer BP-ANNs	Pixel information	167 mammograms with pathologies and 89 without pathologies.; 50% for training and 50% for testing.	Test on 20 out of 128 mammograms covering microcalcification size-range of; 50–250 μm: 0.9% FP at 85% TP,; 2.4% FP at 100% TP.; 50–2,000 μm: 25% FP at 84% TP,; 40% FP at 100% TP.
Zhang et al. (1994) [30]	CADe	The Shift-Invariant ANN (SI-ANN)	Pixel information	168 ROIs from 34 digitized mammograms.; 50%-50% cross-validation.	ROC index: 𝐴𝑍=0.91±0.02,; 45% FP at 100% TP.
Chan et al. (1995) [35]	CADe	The Convolution Neural Network (CNN)	Pixel information	52 mammograms; Group 1: 110 TP and 116 FP.; Group 2: 108 TP and 116 FP.; Two-fold cross-validation.	ROC index: 𝐴𝑍=0.9.; FP rate: 0.1 cluster per image at 100% TP (for obvious cases),; 1.5 cluster per image at 90% TP (for average and subtle cases).
Nagel et al. (1998) [36]	CADe	SI-ANN	Features extracted from image	196 TPs  and  1,252 FPs.; Leave-one-out cross-validation.	The number of FPs per image at 83% TP: 1.6 for ANN,; 0.8 for ANN and rule-based method.; Average ROC index: 𝐴𝑍=0.85 (stdev = 0.04) for ANN,; 𝐴𝑍=0.64 (stdev = 0.07) for ANN + rule-based method (𝑃=0.014).
Wu et al. (1992) [34]	CADe	BP-ANN	Pixel information	56 positive, 56 negative, and 56 FP ROIs, respectively.; 50%-50% cross-validation.	For individual microcalcifications: 𝐴𝑍=0.71.; For clustered microcalcifications: 𝐴𝑍=0.85; 50% FP at 95% TP.
Jiang et al. (1996) [45]	CADx	BP-ANN	Computer-extracted morphological features	40 malignant and 67 benign cases from 100 images.; Leave-one-out cross-validation.	Identified 100% malignant and 82% of the benign cases.; Significantly better than radiologists without computer aid (𝑃=0.03).
Jiang et al. (1999) [46]	CADx	BP-ANN	Computer-extracted morphological features	46 malignant and 58 benign cases.; Leave-one-out cross-validation.	By 10 radiologists: 𝐴𝑍=0.61, sensitivity = 73.5%, specificity = 31.6%.; With aid of ANN: 𝐴𝑍=0.75 (𝑃<0.0001), sensitivity = 87.4%, specificity = 41.9%.
Huo et al. (1998) [47]	CADx	BP-ANN	Morphological features characterizing margin and density	38 benign and 57 malignant cases from 65 patients.; Leave-one-out cross-validation.	ANN: 𝐴𝑍=0.90,; 19.2% specificity at 100% sensitivity.; Hybrid method (rule-based + ANN): 𝐴𝑍=0.94,; 69.2% specificity at 100% sensitivity.
Kallergi (2004) [22]	CADx	BP-ANN	Morphological and distributional descriptors	50 benign and 50 malignant cases.; Leave-one-out cross-validation.	𝐴𝑍=0.98±0.01,; 85% specificity at 100% sensitivity.
Chan et al. (1997) [48]	CADx	BP-ANN	Texture features SGLD matrices	41 malignant and 45 benign cases from 54 patients.; Leave-one-out cross-validation.	With best subset of features: Mammogram-by-mammogram: 𝐴𝑍=0.88,; 24% specificity at 100% sensitivity.; Patient-by-patient: 39% specificity at 100% sensitivity.
Baker et al. (1995) [41]	CADx	BP-ANN	BI-RADS lesion descriptors and medical history variables	133 benign and 73 malignant cases.; Leave-one-out cross-validation.	PPV: 61% (ANN) versus 35% (radiologists).; ROC index: 𝐴𝑍=0.89±0.02 (ANN) versus 0.85±0.03 (radiologists), 𝑃=0.29.; Specificity: 62% (ANN) versus 30% (radiologists) at 100% sensitivity (𝑃<0.01).
Lo et al. (1999) [42]	CADx	BP-ANN	BI-RADS lesion descriptors, age, and history variables	326 benign and 174 malignant cases.; Leave-one-out cross-validation.	Only BI-RADS features: 𝐴𝑍=0.84±0.02, 6% specificity at 100% sensitivity.; BI-RADS + age: 𝐴𝑍=0.86±0.02, 30% specificity at 100% sensitivity.; All features: 𝐴𝑍=0.87±0.02, 22% specificity at 100% sensitivity.; Age variable significantly improves the 𝐴𝑍 (𝑃=0.028).
Ayer et al. (2010) [43]	CADx	BP-ANN	Demographic, mammographic features, and BI-RADS categories	510 malignant and 61,709 benign cases.; 10-fold cross-validation.	𝐴𝑍=0.965 (ANN) versus 0.939 (radiologists), 𝑃<0.001.; Specificity at 85% sensitivity: 94.5% (ANN) versus 88.2% (radiologists), 𝑃<0.001.
Jesneck et al. (2007) [49]	CADx	BP-ANN	Mammographic features, sonographic features, and history features	296 malignant and 507 benign cases.; 500 for training and validation, 303 for testing.	Training and validation set: 𝐴𝑍=0.92±0.01,; Testing set: 𝐴𝑍=0.91±0.02.
Tourassi et al. (2003) [51]	CADx	CSNN	BI-RADS features, age and history	Training set: 174 malignant and 326 benign cases.; Testing set: 358 malignant and 672 benign cases.	On training set: 𝐴𝑍=0.84±0.02; On testing set: 𝐴𝑍=0.81±0.02; CSNN is also capable to impute missing data.
Orr (2001) [54]	CADx and risk estimation	BP-ANN	Age and radiographic features	185 malignant and 1,103 benign cases.; 490 for training and the rest for testing.	𝐴𝑍=0.89 (surgeons) versus 0.86 (ANN), 𝑃=0.004.; ANN is possible for risk stratification.

CADe: computer-aided detection, CADx: computer-aided diagnosis, ANN: artificial neural network, BP-ANN: back-propagation artificial neural network, FP: false positive, TP: true positive, ROI: region of interest, SGLD: spatial grey level dependence, PPV: positive prediction value, BI-RADS: the breast imaging reporting and data system, CSNN: constraint satisfaction neural network, and SI-ANN: shift-invariant artificial neural network.