Ranking-Aware Multiple Instance Learning for Histopathology Slide Classification

Ho Heon Kim, Gisu Hwang, Won Chan Jeong, Young Sin Go

Keywords: Multiple instance learning, learning to rank, digital pathology

TL;DR: We introduce Rank induction, a MIL framework that leverages expert annotations through pairwise ranking between lesion and non-lesion patches to guide attention learning

Abstract: In digital pathology, most deep learning models adopt multiple instance learning (MIL) as it requires only slide-level labels, reducing the need for detailed annotations. However, since MIL still relies on large datasets, data-efficient strategies have emerged as promising alternatives. Although some datasets include expert annotations, their integration with MIL to take advantage of this valuable information has been overlooked. We propose Rank induction, a method that ranks annotated lesion areas against non-lesion areas to guide the model’s attention toward diagnostically meaningful areas. Our experiments on the Camelyon16 dataset show that Rank induction outperforms existing approaches in classification performance. Furthermore, the method remains robust under data-scarce conditions. Finally, attention maps generated by the model trained with Rank induction focus more accurately on cancerous areas.

Submission Number: 41