Sparse concept attribution for histomorphological hypothesis generation from whole-slide classifiers
Abstract
Histology images contain rich morphological information and can provide insights into pathological processes. However, deriving hypotheses relating morphological phenotypes to clinical attributes is bottlenecked by a manual image interpretation step. Here, we demonstrate that this process can be automated through interpretable deep learning. We present SCOPE, a method to interpret slide-level classifiers by combining pathology-specific vision--language models with sparse concept attribution onto a generalist histomorphological concept bank. To measure whether such explanations recover known morphology, we introduce MorphoRecoveryBench, a benchmark of seven tasks with pathologist-curated reference descriptions. On this benchmark, dense concept attribution is indistinguishable from a random baseline, whereas sparse attribution recovers substantial known morphology; decomposing the pooled slide embedding reaches similar explanation correctness at a fraction of the computational cost. Post-hoc interpretation of whole-slide classifiers can thus generate morphological hypotheses at scale, for expert validation.
Related
- Cerebrospinal fluid metagenomic next-generation sequencing for culture-negative anaerobic central nervous system infection involving Prevotella oris: attribution challenges in low-biomass specimens
- An immune world model for multiscale forecasting and therapeutic hypothesis generation
- HypoKG: Evidence-Disciplined Biomedical Hypothesis Generation Beyond Endpoint Knowledge
- Characterizing LLM scientific concept generation: A multi-dimensional measurement study
- General Purpose AI for Public Data Analysis and Experimental Hypothesis Generation
BibTeX
@misc{lazard2026sparse,
title = {{Sparse concept attribution for histomorphological hypothesis generation from whole-slide classifiers}},
author = {Tristan Lazard and Kenza Bouzid and Julius Hense and Shruthi Bannur and Daniel Coelho de Castro and Daniel Shao and Rajesh Jena and Drew Williamson and Stephanie Hyland},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.02985},
eprint = {2609.02985},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.02985},
}
RIS
TY - UNPB
TI - Sparse concept attribution for histomorphological hypothesis generation from whole-slide classifiers
AU - Tristan Lazard
AU - Kenza Bouzid
AU - Julius Hense
AU - Shruthi Bannur
AU - Daniel Coelho de Castro
AU - Daniel Shao
AU - Rajesh Jena
AU - Drew Williamson
AU - Stephanie Hyland
PY - 2026
DA - 2026/09/02
PB - arXiv
DO - 10.48550/arxiv.2609.02985
UR - https://arxiv.org/abs/2609.02985
AB - Histology images contain rich morphological information and can provide insights into pathological processes. However, deriving hypotheses relating morphological phenotypes to clinical attributes is bottlenecked by a manual image interpretation step. Here, we demonstrate that this process can be automated through interpretable deep learning. We present SCOPE, a method to interpret slide-level classifiers by combining pathology-specific vision--language models with sparse concept attribution onto a generalist histomorphological concept bank. To measure whether such explanations recover known morphology, we introduce MorphoRecoveryBench, a benchmark of seven tasks with pathologist-curated reference descriptions. On this benchmark, dense concept attribution is indistinguishable from a random baseline, whereas sparse attribution recovers substantial known morphology; decomposing the pooled slide embedding reaches similar explanation correctness at a fraction of the computational cost. Post-hoc interpretation of whole-slide classifiers can thus generate morphological hypotheses at scale, for expert validation.
ER -