Comparative Study of Deep Learning Models For Automatic Detection of Metastases in H&E Stained Images

Deep learning has gained the attention of researchers around the world. We used several deep learning models to check whether they could potentially be used as a clinical utility by pathologists to detect lymph node metastases in tissue sections of women with breast cancer to save time and reduce the burden on healthcare. For that purpose, we selected three deep learning models of different depths. The deepest one (DenseNet201), medium depth (Resent50), lightweight (Mobie/Net) and trained them in different experiments under various hyperparameters settings from scratch. The test set consists of 32000 histopathological images. We used absolute accuracy, sensitivity, specificity, and F1-score as a performance measure. Apart from these measures, we observe the effect of data augmentation and the number of epochs on the classification performance. We achieved an average accuracy of 85.6%, 87.5%, and 88.4% for MobileNet V1, ResNet50, and DenseNet201, respectively, which further improves to 91.1 %, 93.8%, and 95.7%, respectively, when augmented images are used in the training set. The performance of all deep learning models was on par with that of expert pathologists reported in different studies. We may expect that deep learning could be an exceptional utility for clinicians in real-time diagnosis too.