KMIT-Pathology: Digital Pathology AI Platform for Cancer Biomarkers Identification on Whole Slide Images

Abstract

Analysis and identification of cancer imaging bio markers on biopsy tissues are done through optical microscope. Digital tissue scanners and Deep learning models automate this task and produce unbiased diagnostics. The digital tissue scanner is called as virtual microscopy which digitize the glass slide tissues and the digitized images are called as Whole Slide Images (WSI). They are multi-layered (level) images having high resolution, huge in size and stored as a pyramidal tiff file. As normal web browsers are unable to handle WSI, a special web imaging platform is needed to obtain, store, visualize and process WSI. This platform must provide basic facilities for uploading, viewing and annotating WSI which are the inputs to the deep learning models. The integration of deep learning models with the platform and the WSI database provides a complete solution to cancer diagnostics and detection. This paper proposes two AI deep learning models for the diagnostics and the detection of cancer imaging bio markers on breast cancer and prostate cancer WSI. Efficientnet deep learning model is used to detect ISUP (International Society of Urologic Pathologists) grading for prostate cancer which is trained and tested by 5000 prostate WSI and produces 80% accuracy with 0.6898 quadratic weighted kappa (QWK) score. R2Unet model is used to identify tubule structures for breast cancer which is a morphological component to grade breast cancer. The model is trained and tested by 17432 WSI tiles and generates f1 metric accuracy as 0.9961 with mean_io_u 0.8612. The paper also shows the complete execution of these two Deep learning models (from uploading WSI to visualize the AI detected results) on the newly developed WSI imaging web platform.