Automated Gleason Grading of Whole Slide Images

This repository contains the work done as a part of the R&D Course at Medical Deep Learning and AI Lab at EE Department, IIT Bombay. I have done this project in collaboration with Parth Dodhia and Gautam Kumar under the guidance of Prof. Amit Sethi. You can find the report and presentation respectively.
logo.png

Abstract

Prostrate Cancer is the second leading cause of cancer deaths in men[1]. The Gleason Grading Systems was developed to find the severity of cancer and grade them accordingly based on some specific heterogeneous pattern. This Gleason Grading requires highly trained pathologist. We have designed a automated Annotation system using Deep Learning, where given a WSI(Whole Slide Image) of a Patient, the model predict the type of Gleason Grade. We have trained our system on 641 patients and then evaluated on an independent set of 245 patients. Availability of annotated ground truths enabled us to implement a segmentation model. We also experimented with attention based MIL models on the patch level. We also did not expect a very clear boundary between two different grades of cancer in the same histopathology image, and we were able to predict majority class, that is the type of Gleason Grade reasonably well. Finally, we found that the model performed well in discriminating between Gleason grade 3 and 4.

Overview

Prostatic Carcinomas are graded according to the Gleason scoring system which was first established by Donald Gleason in 1966[2]. The Gleason Grading System is acknowledged by the World Health Organization(WHO) and has been modified and revised in 2005 and 2014 by the International Society of Urological Pathology(ISUP)[3].. Though there was several changes in the clinical diagnosis methods, Gleason grading remains as one of powerful prognostic tool.The diagnosis using Gleason Grading is based on pattern of tumours. The histological patterns are given different grades between 1 to 5, 1 indicating well differentiated and 5 indicating poorly differentiated. Gleason pattern 4 includes fused glands, cribriform and glomeruloid structures and poorly formed glands. Gleason pattern 5 involves poorly differentiated individual cells, sheets of tumour, solid nests, cords and linear arrays as well as comedonecrosis. Gleason Grade-3 and Gleason Grade-4 are usually present in pairs, and in order avoid diagnosis error and provide correct treatment an automated solution would be extremely useful. In this report, we present an approach using a UNet Model to classify different Gleason Grades of Prostrate Cancer. We have used our evaluation metric as Cohen’s Kappa since the test set images were labelled by two pathologists and it also has a class imbalance problem. We also experimented with the recent technique of attention based multiple instance learning.

Conclusion

The segmentation approach gave reasonable results with good performance on the test as compared to agreement between pathologists. It did quite well in discriminating between Gleason grades 3 and 4 which is an important factor for diagnosis. Future work could be to feed a multi-resolution input to the UNet for better performance.

Attention based classification was not working for all the classes. The model is easily able to identify Gleason 3 and Gleason 4. The training process was very unstable. We tried various schedulers and increased the depth of the attention and classification layers, but we couldn’t get good results. The model was overfitting the train set with train accuracies reaching more than 80%. Further work can be to use segmentation with self guided attention as mentioned in the given paper.