Bilkent University
Department of Computer Engineering


Transformer Based Methods for Classification of Histopathology Images


Cihan Erkan
Master Student
(Supervisor: Prof.Dr.Selim Aksoy)
Computer Engineering Department
Bilkent University

Abstract: Pathologists can diagnose cancer based on the biopsy samples of the patients as these samples contain the necessary information regarding the existence/severity of the cancer. It is possible to scan these samples to obtain digital images, called Whole Slide Images (WSIs). After digitization of the samples, cancer diagnosis can be formulated as an image classification problem. Unfortunately, widely used image classification methods cannot be directly applied to WSIs due to their shape, size and composition. The most common approach is dividing the WSI into smaller patches and aggregating the patch information to obtain a representation for the WSI. Here, the aggregation step is still an open research direction. Recent works suggest that the transformers are very capable of encoding sequence-like data, which is similar to the patches in a WSI. Because of that, we propose that transformer based architectures can be a promising way of aggregating the patch information. In this presentation, we discuss possible ways of incorporating transformer based methods to the WSI classification pipeline.


DATE: 28 March, Monday @ 15:50