Table 3(A).

Studies using deep learning approach for microcalcification detection and classification in mammography and digital breast tomosynthesis

Journal article	Year	Training set	Validation set	Independenttest set	Convolutional neural network (CNN) structure	Performance* (validation or independent test)
Microcalcification detection
Samala et al.31	2014	78 DBT vols with MC clus (DBT:21PVs, 60o scan)	49 DBT vols with MC clus	104 DBT vols with MC clus, 76 no MC	CNN with two convolution layers	FROC: 85% sens. at 0.71 FP/vol. (view-based), at 0.54 FP/vol (case-based)
Samala et al.32	2015	78 DBT vols with MC clus (DBT:11PVs, 30o scan)	49 DBT vols with MC clus	104 DBT vols with MC clus, 76 no MC	CNN with two convolution layers	FROC 85% sens. at 1.72 FP/vol. (view-based), at 0.49 FP/vol (case-based)
Wang et al.33	2018	167 cases (300 images)	67 cases (117 images)	158 cases (292 images)	Context-sensitive DNN: 7-conv-layer global CNN and 3-conv-layer local CNN (indiv MC 9 × 9, clus 95 × 95 ROIs) compared to clus-based CNN	FROC cluster-based 85% sens: DCNN with 10 conv layers 0.40 FPI; cluster-based CNN 0.44 FPI; SVM 0.52 FPI
Microcalcification classification
Wang et al.34	2016	1000 images (677B, 323M); 10-fold CV		204 images (97B, 107M): 110 MC, 35 mass, 59 both	Stacked autoencoder (SAE) as feature extractor. SVM feature classifier	AUC(MC)=0.87, AUC(mass)=0.61, AUC(MC&mass)=0.90
Cai et al.35	2019	891 images (486M, 405B); 10-fold CV		99 images (54M, 45B)	Fine-tuning of ImageNet-pretrained AlexNet as deep feature extractor. SVM classification of deep features with and without handcrafted features	AUC(M vs B)=0.93–0.94
Shi et al36	2018	99 mag DMs DCIS (25 upstaged to invasive): 80% training, 20% validation			ImageNet-pretrained VGG16 as feature extractor, logistic regression classifier with feature selection	AUC (DCIS vs-upstaged)=0.70