An Approach for Clustering Subjects According to Similarities in Cell Distributions within Biopsies

TL;DR

This paper presents a new interpretable method for clustering cancer patients based on cell distribution patterns in biopsies, using histograms to capture complex cellular repartitions and relate them to prognosis.

Contribution

The study introduces a novel workflow that leverages cell distribution histograms for patient clustering, providing a more detailed analysis of tumor heterogeneity compared to prior methods.

Findings

Clusters correlate with known prognosis indicators.

Method achieves high confidence in matching existing prognosis knowledge.

Workflow is applicable to H&E-stained tissue images.

Abstract

In this paper, we introduce a novel and interpretable methodology to cluster subjects suffering from cancer, based on features extracted from their biopsies. Contrary to existing approaches, we propose here to capture complex patterns in the repartitions of their cells using histograms, and compare subjects on the basis of these repartitions. We describe here our complete workflow, including creation of the database, cells segmentation and phenotyping, computation of complex features, choice of a distance function between features, clustering between subjects using that distance, and survival analysis of obtained clusters. We illustrate our approach on a database of hematoxylin and eosin (H&E)-stained tissues of subjects suffering from Stage I lung adenocarcinoma, where our results match existing knowledge in prognosis estimation with high confidence.