A dataset and a methodology for intraoperative computer-aided diagnosis of a metastatic colon cancer in a liver

The lack of pixel-wise annotated images severely hinders the deep learning approach to computer-aided diagnosis in histopathology. This research creates a public database comprised of: (i) a dataset of 82 histopathological images of hematoxylin-eosin stained frozen sections acquired intraoperatively on 19 patients diagnosed with metastatic colon cancer in a liver; (ii) corresponding pixel-wise ground truth maps annotated by four pathologists, two residents in pathology, and one final-year student of medicine. The Fleiss' kappa equal to 0.74 indicates substantial inter-annotator agreement; (iii) two datasets with images stain-normalized relative to two target images; (iv) development of two conventional machine learning and three deep learning-based diagnostic models. The database is available at http://cocahis.irb.hr. For binary, cancer vs. non-cancer, pixel-wise diagnosis we develop: SVM, kNN, U-Net, U-Net++, and DeepLabv3 classifiers that combine results from original images and stain-normalized images, which can be interpreted as different views. On average, deep learning classifiers outperformed SVM and kNN classifiers on an independent test set 14% in terms of micro balanced accuracy, 15% in terms of the micro F-1 score, and 26% in terms of micro precision. As opposed to that, the difference in performance between deep classifiers is within 2%. We found an insignificant difference in performance between deep classifiers trained from scratch and corresponding classifiers pre-trained on non-domain image datasets. The best micro balanced accuracy estimated on the independent test set by the U-Net++ classifier equals 89.34%. Corresponding amounts of F-1 score and precision are, respectively, 83.67% and 81.11%.