Enhancing Cell Detection with Transformer-Based Architectures in Multi-Level Magnification Classification for Computational Pathology

Abstract

Cell detection and classification are important tasks in aiding patient prognosis and treatment planning in Computational Pathology (CPATH). Pathologists usually consider different levels of magnification when making diagnoses. Inspired by this, recent methods in Machine Learning (ML) have been proposed to utilize the cell-tissue relationship with different levels of magnification when detecting and classifying cells. In particular, a new dataset named OCELOT was released, containing overlapping cell and tissue annotations based on Hematoxylin and Eosin (H&E) stained Whole Slide Images (WSIs) of multiple organs. Although good results were reached on the OCELOT dataset initially, they were all limited to models based on Convolutional Neural Networks (CNNs) that were years behind the state-of-the-art in Computer Vision (CV) today. The OCELOT dataset was posted as a challenge online, yielding submissions with newer architectures. In this work, we explore the use of transformer-based architecture on the OCELOT dataset and propose a new model architecture specifically made to leverage the added tissue context, which reaches state-of-the-art performance with an F1 score of 72.62% on the official OCELOT test set. Additionally, we explore how the tissue context is used by the models.