Self-Supervision for Renal Cell Carcinoma Subtyping

Cancer is a disease that occurs when abnormal cells grow and spread uncontrollably in the body. There are various types of cancer, each with their own unique characteristics and treatment options. Renal cell carcinoma (RCC) is a form of kidney cancer that can be classified into several histological subtypes. RCC is the most common type of kidney cancer in adults, making up roughly 90% of cases. Accurately identifying the tumor subtype is critical as treatment approaches and prognosis may differ based on the subtype and stage of the disease. Some of the major subtypes of RCC include clear cell carcinoma (ccRCC), papillary (pRCC), chromophobe, oncocytoma, and others. Deep learning models, specifically convolutional neural networks (CNNs), have shown great promise in the medical domain for classifying, segmenting, and detecting objects in images. However, the medical field poses unique challenges due to the lack of large and diverse datasets like ImageNet. Self-supervised techniques have been developed to address this issue, but most proposed methods for histopathological images do not fully exploit the domain data property. While they utilize the different magnifications present in whole slide images, they miss the fact that those various fields of view are interconnected. To address this gap, we propose a new self-supervised task that aims to interconnect different magnification levels of histopathological images. Our proposed task requires the model to localize a tile inside a global patch, the image representing the tile at a higher resolution is extracted from a magnification level higher than that of the global patch. Both the higher-resolution image and the global patch are given to the network to perform the localization pre-text task. We present different possible formulations for this problem and compare our method to a newly introduced state-of-the-art pretext task. Our results show that our method performs on par or better than its counterpart and outperforms models pre-trained on ImageNet. We also discuss the challenges faced during training and how we handle creating and processing our dataset. Finally, for our future work, we highlight the potential of our proposed method in a reinforcement learning framework for the efficient processing of whole slide images. Our approach has the potential to significantly improve the accuracy of histopathological image analysis, which could lead to better diagnosis, prognosis, and treatment of cancer.