CLEAR-HPV: Interpretable Concept Discovery for HPV-Associated Morphology in Whole-Slide Histology

Human papillomavirus (HPV) status is a critical determinant of prognosis and treatment response in head and neck and cervical cancers. Although attention-based multiple instance learning (MIL) achieves strong slide-level prediction for HPV-related whole-slide histopathology, it provides limited morphologic interpretability. To address this limitation, we introduce Concept-Level Explainable Attention-guided Representation for HPV (CLEAR-HPV), a framework that restructures the MIL latent space using attention to enable concept discovery without requiring concept labels during training. Operating in an attention-weighted latent space, CLEAR-HPV automatically discovers keratinizing, basaloid, and stromal morphologic concepts, generates spatial concept maps, and represents each slide using a compact concept-fraction vector. CLEAR-HPV’s concept-fraction vectors preserve the predictive information of the original MIL embeddings while reducing the high-dimensional feature space (e.g., 1536 dimensions) to only 10 interpretable concepts. CLEAR-HPV generalizes consistently across TCGA-HNSCC, TCGA-CESC, and CPTAC-HNSCC, providing compact, concept-level interpretability through a general, backbone-agnostic framework for attention-based MIL models of whole-slide histopathology.

💡 Research Summary

The paper introduces CLEAR‑HPV (Concept‑Level Explainable Attention‑guided Representation for HPV), a post‑hoc framework that extracts interpretable histopathologic concepts from attention‑based multiple instance learning (MIL) models used for whole‑slide image (WSI) classification of HPV status in head‑and‑neck and cervical cancers. Traditional attention‑based MIL methods such as CLAM achieve strong slide‑level performance but offer limited insight into which morphological patterns drive predictions. CLEAR‑HPV addresses this gap by re‑weighting the tile‑level latent embeddings (the “h‑space”) with the attention scores learned by the MIL backbone, thereby emphasizing diagnostically informative regions while suppressing background variability.

In the attention‑weighted h‑space, the method applies K‑means clustering (K=10, chosen via elbow analysis) to discover discrete, annotation‑free concepts. The resulting clusters correspond to biologically meaningful morphologies—keratinizing, basaloid, and stromal patterns—that align with known HPV‑associated histology. Each slide is then summarized by a 10‑dimensional concept‑fraction vector, representing the proportion of tiles assigned to each concept. This compact representation retains the predictive signal of the original high‑dimensional (e.g., 1536‑dim) MIL embeddings, enabling a simple, parameter‑free concept‑fraction classifier to achieve performance comparable to the original model (e.g., AUC ≈ 0.84 vs. 0.86 for CLAM).

The authors evaluate CLEAR‑HPV on three independent cohorts: TCGA‑HNSCC (102 patients, 38 HPV‑positive), TCGA‑CESC (146 patients, 138 HPV‑positive), and CPTAC‑HNSCC (112 HPV‑negative patients). Across these datasets, CLEAR‑HPV consistently discovers stable concepts and maintains robust slide‑level accuracy, even in zero‑shot settings without any retraining. Comparative experiments with baseline concept‑discovery approaches—heatmap‑only grouping, encoder‑feature clustering, and a Dirichlet mixture model—show that only the attention‑guided method preserves both interpretability and predictive power.

Key contributions include: (1) the first general framework for automatic, annotation‑free discovery of pathology‑relevant concepts in HPV prediction; (2) a dual output of spatial concept maps and low‑dimensional concept‑fraction vectors that provide biologically grounded explanations; (3) dimensionality reduction from thousands of features to ten interpretable concepts without loss of classification performance; (4) backbone‑agnostic applicability, demonstrated across multiple MIL architectures; and (5) cross‑cohort stability, with concept‑fraction representations sometimes outperforming the original MIL models.

Limitations noted are the need to pre‑select the number of concepts, sensitivity of clustering to initialization, and the focus on a single molecular marker (HPV). Future work will explore automated K selection, extension to other biomarkers and multi‑label tasks, and validation of discovered concepts through pathologist feedback.

Overall, CLEAR‑HPV reveals that attention‑based MIL models already encode rich morphological structure, and that a simple attention‑guided re‑organization can make this structure explicit, delivering both high diagnostic accuracy and clinically meaningful interpretability for digital pathology.