Context Matters: Anatomy-Aware Dual-Stream Multiple Instance Learning Framework for eGFR Prediction

Abteen Arab, Maziar Riazy, Ali Bashashati

Keywords: Kidney Disease, Computational Pathology, Multiple Instance Learning (MIL), Representation Learning

TL;DR: We show that kidney WSIs contain disease-dependent, function-related signals and that modeling glomerular/non-glomerular tissue separately improves eGFR prediction over standard single-stream MIL.

Registration Requirement: Yes

Abstract: Current deep learning approaches for modeling kidney disease typically rely on pathologists' annotations, which are costly to obtain and subject to inter-observer variability. In contrast, estimated glomerular filtration rate (eGFR) is a routine clinical measure of kidney function that offers a more scalable supervision signal, yet remains relatively underexplored. In this work, we assess the performance of multiple instance learning (MIL) frameworks for eGFR prediction from kidney whole slide images (WSIs). Importantly, we show that the predictive value of different tissue segments is disease dependent and that modelling kidney disease as two anatomy-defined MIL streams yields stronger performance than standard single stream baselines.

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 26