| Table 3(C). Studies using deep learning approach for breast density segmentation and classification in mammography. | ||||||
|---|---|---|---|---|---|---|
| Journal article | Year | Training set | Validation set | Independenttest set | CNN structure | Performance (validation or independent test) |
| Breast density segmentation | ||||||
| Kallenberg et al.64 | 2016 | Set1: 493N views; Set2: (226 cases, 442 controls); Set3: (394 cases, 1182 controls); 5-fold CV | Convolutional sparse autoencoder (CSAE). (1) density segmentation (MD); (2) case-vs-control classification (MT) | Correlation coeff. (MD)=0.85; Set3: AUC(MT-CSAE)=0.57; AUC(MT-density)=0.59 | ||
| Li et al.65 | 2018 | 478 DMs; 10-fold CV | 183 DMs | DCNN with three convolutional layers | DSC = 0.76; Correl coeff. = 0.94 | |
| Mohamed et al.66 | 2018 | BI-RADS density B and C: 7000 DMs each; 6-fold CV | BI-RADS density B and C: 925 images each | Modified AlexNet: ImageNet-pretrained vs training from scratch. | AUC(scratch)=0.94, AUC(pretrained)=0.92 | |
| Mohamed et al.67 | 2018 | 963N cases with 15,415 DMs. BI-RADS density from clinical reports; 70% training, 15% validation | 15% | Modified AlexNets for two tasks: (1) BI-RADS B-vs-C, (2) Dense (A&B)-vs-nondense (C&D) | (1) AUC (CC&MLO)=0.92; (2) AUC(CC&MLO)=0.95 | |
| Lee et al.68 | 2018 | 455 DM cases | 58 DM cases | 91 DM cases | ImageNet-pretrained VGG16 | Correl coeff. % density-vs-BI-RADS (radiologist): CC=0.81, MLO=0.79, average=0.85 |
| Wanders et al.69 | 2018 | 394 cancers, 1182 controls (DMs) | 51,400 (301 cancer, 51,099 controls; DMs) | DCNN by Kallenberg et al.64 | C-index: Texture + vol density = 0.62, vol density = 0.56 | |
| Gastounioti et al.70 | 2018 | 200 pts (1:3 case:control; DMs) | 100 pts (1:3 case:control; DMs) | 124 pts (1:3 case:control; DMs) | LeNet-like CNN with 29 input channels with texture images, two convolutional layers. DCNN with DM input, five convolutional layers | Case-vs-control: AUC(DCNN-multichannel texture) = 0.90, AUC(DCNN-DM) = 0.63 |
| Ciritsis et al.71 | 2018 | 70% of 12,392 views (6470 RMLO, 6462 RCC) | 30% of 12,392 | Set 1: (850 MLO, 882 CC); Set 2: (100 MLO, 100 CC, 2 radiologists' consensus) | DCNN with 13 convolutional layers, four dense layers, output 4 BI-RADS density (A, B, C, D) | Accuracy: Set 1: BI-RADS: 71.0%-71.7%; dense-vs-nondense: 88.6%-89.9%; Set 2: BI-RADS 87.4%-92.2%; dense-vs-nondense 96%-99% |
| Lehman et al.72 | 2019 | 27684 cases (41,479 DMs) | 8738 DMs | 5741 cases (8677 DMs); Clinic test: 10,763 cases | ImageNet-pretrained ResNet18 | Test set BI-RADS: Accuracy=77%, kappa=0.67; Dense-vs-nondense: 87%. Clinic test: BI-RADS: Accuracy = 90%, kappa=0.85 |
AD, architectural distortion; AUC(condition), area under the receiver operating characteristic (ROC) curve for the condition in parenthesis; B, benign; CC, craniocaudal view; CEDM, contrast-enhanced digital mammogram; CESM, contrast-enhanced spectral mammogram; CV, cross validation; DBT, digital breast tomosynthesis; DCIS, ductal carcinoma in situ; DM, digital mammography; DSC, Dice similarity coefficient; FPI, false positives/image; FROC, free-response ROC curve; LSTM, long short-term memory; M, malignant; MC, microcalcification; MLO, mediolateral oblique view; N, normal; SFM, screen-film mammogram; SVM, support vector machine; clus, cluster; indiv, individual; pts, patients; vol, volume.
ImageNet: training data set containing over 1.2 million photographic images from the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) for classification of over 1000 classes of everyday objects (cars, animals, planes, etc)