Prototype‑Guided Hyperbolic Multi‑Scale Learning and Density‑Ratio Analysis for Subtype‑Specific Nuclei Discovery in WSIs

Kazumasa Ohara, Kei Taguchi, Ryoichi Koga, Tatsuya Yokota, Hiroaki Miyoshi, Noriaki Hashimoto, Ichiro Takeuchi, Hidekata Hontani

Keywords: Hyperbolic Representation Learning, Pathology Image Analysis, Multi-Scale Representation

Abstract:

We present a self-supervised framework that discovers subtype-specific nuclei directly from whole-slide images (WSIs) by learning a unified, multi-scale representation in a single hyperbolic space.
Lymphoma progression manifests as coordinated morphological changes across scales-from individual nuclei to tissue architecture-yet most existing pipelines encode these scales with separate models and rely on tissue-level supervision, which obscures cell-level drivers of subtype identity.
Our approach replaces this separation with inclusion-aware self-supervision: local crops (nucleus patches) and their containing global crops (tissue patches) are jointly embedded in a Poincaré ball, whose hyperbolic geometry naturally accommodates hierarchical structure.
A cross-scale alignment objective pulls each nucleus toward the representation of its parent tissue region, enabling a single encoder to capture fine-to-global morphology without cell-level labels.
Tissue patches sampled from lesions typically carry features sufficient for subtype discrimination and form clusters reflecting both lesional status and subtype.
By contrast, single-nucleus patches are often weakly informative in isolation and inter-subtype differences at the tissue level largely arise from composition ratios of common nuclear phenotypes.
This approach yields nucleus-level exemplars and spatial patterns that are consistent with tissue-level subtype structure while preserving interpretability at cellular resolution.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 24315