Miljana Shulajkovska, Matej Jelenc, Jitenndra Jonnagaddala and Anton Gradišek

Abstract
Microsatellite instability (MSI) is a crucial biomarker in colorectal cancer, guiding personalised treatment strategies. The focus of our paper is on evaluating how different state-of-the-art pretrained artificial intelligence models perform in extracting features on molecular and cellular oncology (MCO) study dataset to predict biomarkers. In this study, we present an advanced approach for MSI prediction using multiple instance learning on whole slide images. Our process begins with comprehensive preprocessing of WSIs, followed by tessellation, which breaks down large images into manageable tiles. State-of-the-art feature extraction techniques are utilised on these selected tiles, employing pretrained models to capture rich, discriminative features. Various aggregation methods are applied to combine these features,
leading to the prediction of MSI status across the entire slide. We assess the performance of different pretrained models within this framework, demonstrating their effectiveness in accurately predicting MSI, with results showing an AUROC of 0.91 on the MCO dataset. Our findings underscore the potential of multiple instance learning-based approaches in enhancing biomarker prediction in colorectal cancer, contributing to more targeted and effective treatment strategies.