Abstract
Multiple instance learning (MIL) is a promising approach for weaklysupervised classification in pathology using whole slide images (WSIs).However, conventional MIL methods such as Attention-Based Deep MultipleInstance Learning (ABMIL) typically disregard spatial interactions amongpatches that are crucial to pathological diagnosis. Recent advancements, suchas Transformer based MIL (TransMIL), have incorporated spatial context andinter-patch relationships. However, it remains unclear whether explicitlymodeling patch relationships yields similar performance gains in ABMIL, whichrelies solely on Multi-Layer Perceptrons (MLPs). In contrast, TransMIL employsTransformer-based layers, introducing a fundamental architectural shift at thecost of substantially increased computational complexity. In this work, weenhance the ABMIL framework by integrating interaction-aware representations toaddress this question. Our proposed model, Global ABMIL (GABMIL), explicitlycaptures inter-instance dependencies while preserving computational efficiency.Experimental results on two publicly available datasets for tumor subtyping inbreast and lung cancers demonstrate that GABMIL achieves up to a 7 percentagepoint improvement in AUPRC and a 5 percentage point increase in the Kappa scoreover ABMIL, with minimal or no additional computational overhead. Thesefindings underscore the importance of incorporating patch interactions withinMIL frameworks. Our code is available at\href{https://github.com/tueimage/GABMIL}{\texttt{GABMIL}}.