Automated gleason grading of prostate cancer using semi-supervised techniques

• Main Author: Cheong, Isaac Jiasheng
• Other Authors: Lin Guosheng
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2020
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/138143
• Institution: Nanyang Technological University

The Gleason grading system is the most widely used system for determining the aggressiveness of prostate cancer in patients. Using such a system to grade prostate cancer is not only tedious and time-consuming, but it also requires the expertise of trained pathologists. Furthermore, such annotations even suffer from both intra and inter-pathologist variability due to the subjective nature of grading prostate cancer. Given all these issues with manually grading prostate cancer, the quality of treatment of patients would, ultimately, be adversely affected. As a result, using convolutional neural networks (CNNs) to automate this task will help to improve consistency by getting rid of any form of intra-pathologist variability and reduce the workload of pathologists, thereby allowing them to devote more of their time to the more ambiguous cases. Given the aforementioned difficulties with obtaining annotated prostate biopsy images, it is difficult to train a robust model that is able to accurately segment and grade prostate cancer, and generalize to different populations, races, ages, staining procedures, etc. In this study, I explore a new approach to tackling the problem of automated Gleason grading through the use of a semi-supervised learning approach, known as the noisy student method. By using such an approach, we will be able to leverage on unlabeled prostate whole-slide images (WSIs) to train our model and not just rely on the limited amounts of annotated datasets. Furthermore, given that our model can be trained on unlabeled data, we can add prostate tissue images from different populations, or healthcare institutions to our dataset which allows our model to better generalize to different scenarios and thus we will be able to produce a model that is more robust. Although, I have not been able to beat the current state-of-the-art results with this method, I believe that upon performing further experiments, this method should be able to yield desirable results.