Equivariant Neural Field based Whole-Slide representations for microsatellite instability prediction

Marga J. Don, David R Wessels, Ylva A. Weeda, Sybren L. Meijer, Hoel Kervadec, Erik J Bekkers

Keywords: Equivariance, Neural Fields, Whole-Slide Classification, Continuous Representations

Abstract: Predicting microsatellite instability (MSI) from whole-slide images (WSIs) is crucial for immunotherapy patient selection. While Deep Learning has succeeded in Colorectal Cancer (CRC), performance in Gastric Adenocarcinoma (GA) remains limited due to the subtle, less discriminative morphological features associated with MSI in gastric tissue. Existing Multiple Instance Learning (MIL) approaches typically rely on global patch descriptors, often failing to capture the fine-grained local geometric structures required for this task. In this paper, we propose a novel framework utilizing Equivariant Neural Fields (ENFs) for histopathology representation learning. Unlike conventional neural fields that compress signals into a single global latent vector, ENFs represent tissue patches as latent point clouds, explicitly grounding representations in local geometry and ensuring equivariance to rotation. We further propose a hierarchical pipeline where these patch-level point clouds are stitched to form a comprehensive WSI-level representation, which is processed by the Erwin architecture for slide-level prediction. We validate our method on the NCT-CRC-HE-100K dataset and a clinical GA cohort. Our experiments demonstrate that ENFs achieve superior reconstruction fidelity compared to non-geometric Neural Field baselines (MedFuncta) and produce highly informative representations that improve downstream MSI classification performance in challenging gastric adenocarcinoma cases.

Primary Subject Area: Geometric Deep Learning

Secondary Subject Area: Unsupervised Learning and Representation Learning

Registration Requirement: Yes

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 294