Table 5.

Studies using deep learning approach for breast MRI

Journal article	Year	Training set	Validation set	Independent test set	Convolutional neural network (CNN) structure	Performance (validation or independent test)
Classification
Rasti et al.88	2017	112 pts (53M, 59B); 5-fold CV			CNN (three convolutional layers); mixture ensemble of CNNs (ME-CNN) with three CNNs and a convolutional gating network	AUC(CNN)=0.95 AUC(ME-CNN)=0.99
Antropova et al.44	2017	690 pts (478M, 212B); 5-fold CV			ImageNet-pretrained VGG19 as feature extractor, SVM classifier	AUC(Maxpool features)=0.87; AUC(fused with radiomic features)=0.89
Antropova et al.89	2018	690 pts (478M, 212B); maximum intensity projection (MIP), 5-fold CV			ImageNet-pretrained VGG19 as feature extractor, SVM classifier	AUC(MIP features)=0.88
Segmentation/Classification
Zhang et al.90	2019	224 pts (combination of CV and using all data for training different U-Nets)		48 pts	Multiple U-Nets for breast segmentation, landmark detection and mass segmentation	DSC(mass segment.)=71.8 AUC(Luminal A vs others)=0.69 for variance of time-to-peak kinetic feature
Truhn et al.91	2019	447 pts (787M from 341 pts, 507B from 237 pts); 10-fold CV in outer loop and 5-fold CV in innerloop			CNN: ImageNet-pretrained ResNet18 Radiomics: PCA and L1 regularization 562 radiomics	AUC(CNN)=0.88 AUC(radiomics-PCA)=0.78 AUC(radiomics-L1 regularization)=0.81 AUC(radiologist)=0.98
Segmentation of fibroglandular tissue/breast density assessment
Dalmış et al.92	2017	39 pts	5 pts	22 pts	3-class U-Net (non-breast, fatty tissue, fibroglandular tissue)	DSC(breast)=0.93; DSC(FGT)=0.85; Correlation(manual vs U-Net segmented FGT)=0.97

PCA, principal component analysis.