How it started?
According to global cancer statistics, stomach and colon cancers are amongst the most common leading causes of cancer deaths in the world, with stomach cancer ranking fourth in men and seventh in women, and colon cancer ranking third in men and second in women. Histopathological classification of gastric and colonic epithelial tumours is one of the routine pathological diagnosis tasks for pathologists. Computational pathology techniques based on Artificial intelligence (AI) would be of high benefit in easing the ever increasing workloads on pathologists, especially in regions that have shortages in access to pathological diagnosis services. Our latest results demonstrate the generalisation ability of our models and the high promising potential of deployment in a practical histopathological diagnostic workflow system.
Hypothesis and Challenges
We set out to test the hypothesis that it is possible to train generalisable deep learning models for the classification of epithelial tumours in biopsy whole slide images (WSIs) of stomach and colon that achieve a performance similar to human pathologists. Deep learning requires large amounts of good quality annotated images, and one of the main challenges was acquiring and annotating a large dataset for such a purpose. We ended up collecting and annotating 4,128 stomach and 4,036 colon WSIs; however, proprietary tools to improve the efficiency of annotations significantly reduced the working time required for annotation.
Results & Discussion
One of the primary concerns of applying deep learning models is whether they generalise to new images obtained from different medical institutions. In this study, we evaluated each of the models on three test sets of roughly 500 WSI each obtained from different sources. The models on three independent test sets each demonstrated high AUC performances up to 0.97 and 0.99 for gastric adenocarcinoma and adenoma, respectively, and 0.96 and 0.99 for colonic adenocarcinoma and adenoma respectively. In a 30-second time constrained evaluation of the stomach model against a group of pathologists, the model achieved as good as the best pathologist in the group. Though pathologists are certainly under large workloads, they do not usually work under such unrealistic time-constraints; however, this does show the primary advantage of using deep learning as part of the pathologists’ clinical workflow: the speed at which it can operate. It could allow for a much faster turnaround for pathologists by allowing them to focus on the most potentially urgent cases.
Now, we are developing models to classify carcinoma subtypes (e.g., papillary adenocarcinoma, tubular adenocarcinoma, mucinous adenocarcinoma, signet-ring cell carcinoma, adenosquamous carcinoma, etc), which play an important role in developing further practical aid for surgical pathologists, and to train models to predict patient outcome. We would like to witness a future where pathologists and AI collaborate together to support the foundation of medical care.
Fahdi Kanavati: Medmain Research, Medmain Inc.
Masayuki Tsuneki: Board Director & VP of Medical Research, Medmain Research, Medmain Inc.
Iizuka O, Kanavati F, Kato K, Rambeau M, Arihiro K, Tsuneki M. Deep Learning Models for Histopathological Classification of Gastric and Colonic Epithelial Tumours. Sci Rep 10, 1504 (2020). https://doi.org/10.1038/s41598-020-58467-9
Digital Pathology AI Analysis Solution - PidPort