Automatic Prostate Cancer Grading Using Deep Architectures

Abstract

Prostate cancer is the second most aggressive type of cancer diagnosed in men, seriously affecting people’s life and health. Prostate cancer detection and grading in advance is a very critical step for pathologists. Large scale inter observer reproducibility exists in staging the prostate biopsies which leads us to move towards a Computer based model, which could accurately detect and grade the cancerous prostate. Due to recent development in the field of digital pathology, tissue microarrays (TMA) images are generated from whole slide images resulting in less computational procedures and achieve good performance. This thesis is focused on deep learning model to automatically stage the cancer instead of feature engineering based models Deep learning models directly learn the features via convolutional layers and achieve good accuracy as compared to feature engineering based models. We have used two datasets, Harvard dataset and Gleason Challenge 2019, for implementation of our proposed model. Our proposed UNET based architecture is used for training as well as testing and evaluation. We have used different deep learning models for our UNET based encoder and achieved 0.728 and 0.732 average Cohen’s kappa with F1 on both datasets respectively. The results show that our proposed UNET based deep learning model performs better as compared to other state of art models. Hence, it would help pathologists to automatically grade the prostate biopsies with high accuracy.