- Abstract: We propose an automatic method to detect gastric cancer on whole slide images. To do this, we collected cases with slide-level and region-level labels and trained deep neural network for tissue classification. To improve the tissue classification performance, we exploited slide-level weak label for training the model with patches without region-level label. For whole slide classification, we extracted features representative for whole slide characteristics. The proposed method archived 92.51% accuracy on 3 class classification. Macro-average and microaverage AUROC on the test set is 0.9802 and 0.9788.
- Keywords: Digital pathology, Gastric cancer, Deep learning, Medical imaging, Weak Label
- Author Affiliation: Vuno Inc., Green Cross Laboratories