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

TL;DR

CLEAR-HPV reinterprets the attention-guided MIL latent space for HPV-associated histology by discovering discrete, interpretable morphologic concepts without concept-level labels. It operates in the attention-weighted $h$-space to identify keratinizing, basaloid, and stromal patterns, producing spatial concept maps and compact $K$-dimensional concept-fraction vectors (with $K=10$) that preserve the original predictive power. The framework proves backbone-agnostic, generalizes across TCGA-HNSCC, TCGA-CESC, and CPTAC-HNSCC, and demonstrates stable, clinically meaningful concepts across cohorts, enabling interpretability without sacrificing performance. This approach offers a practical path toward interpretable, translatable computational pathology tools and can extend to other MIL-based histopathology tasks.

Abstract

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.

Figures (6)

• Figure 1: Overview of the CLEAR-HPV framework.(A) Data processing pipeline: WSIs are decomposed into fixed-size tiles, encoded with a pretrained ViT or CNN, and converted into patch-level feature embeddings. (B) An attention-based MIL classifier projects embeddings into the $h$-space latent representation and uses multi-head attention to compute tile-level contributions, which are pooled into a single slide-level embedding for HPV prediction. (C) CLEAR-HPV performs annotation-free concept discovery on attention-weighted $h$-space embeddings to identify coherent morphologic concepts. (D) Using the discovered concepts, each slide is represented by a concept-fraction vector, which is then averaged across slides to obtain class-averaged concept-fraction vectors that summarize morphologic composition for HPV-positive and HPV-negative cohorts. (E) Representative tiles illustrate the characteristic morphology captured by each discovered concept. (F) Spatial concept maps visualize the distribution of concepts across the WSIs, revealing their spatial organization. (G) Slide-level concept-fraction vectors provide an interpretable representation that supports a concept-fraction classifier, which recovers MIL predictive performance while offering concept-level explanations. More details are available in the Methods section.
• Figure 2: Recovery score relative to the interpreted MIL model (CLAM) across ACC, AUC, F1, Precision, Recall (i.e., sensitivity), and Specificity. For each method, the Euclidean distance $d$ between its metric vector (i.e., concatenation of Accuracy, AUC, etc.) and the interpreted model's is computed and converted to a similarity score $s = \tfrac{1}{1 + d}$. Higher scores indicate closer agreement with CLAM.
• Figure 3: Class-averaged concept-fraction vectors across concept-discovery settings on TCGA-HNSCC. Concept-fraction vectors are computed per slide as the fraction of tiles assigned to each discovered concept, optionally weighted by MIL attention. These slide-level vectors are then averaged within each group to obtain class-averaged profiles that summarize cohort-level morphologic composition and highlight differences in concept prevalence across clinical groups. Panels (A--C) show group-averaged profiles for HPV-positive (blue) and HPV-negative (orange) cases, while panel (D) shows the corresponding averages for surviving (blue) and deceased (orange) cases. (A) Class-averaged concept-fraction vectors derived from encoder-feature clustering (non-$h$-space baseline). (B) Class-averaged concept-fraction vectors derived from CLEAR-HPV concepts in the MIL $h$-space using unweighted fractions, where all tiles contribute equally. (C) Class-averaged concept-fraction vectors derived from CLEAR-HPV concepts in the MIL $h$-space using attention-weighted fractions, where each tile contributes proportionally to its MIL attention score, yielding clearer separation between HPV-positive and HPV-negative cases. (D) Using the same CLEAR-HPV $h$-space concepts and attention-weighted fractions as in (C), class-averaged concept-fraction vectors are shown for surviving versus deceased cases.
• Figure 4: Top tiles for key concepts discovered by CLEAR-HPV (A) and the corresponding slide-level distributions in the dataset TCGA-HNSCC (B).(A) Top (representative) tiles for five CLEAR-HPV concepts chosen for their consistent appearance and clear morphologic identity: C5 (basaloid squamous epithelium), C7 (keratinizing squamous epithelium), C9 (fibrous stroma), C4 (connective stroma), and C2 (inflammatory cells). (B) Average concept-fraction vectors for HPV-positive (blue) and HPV-negative (orange) slides, with 95% bootstrap confidence intervals. The "Basaloid" Concept C5 is more prevalent in HPV-positive cases, while the "keratinizing" Concept C7 is more prevalent in HPV-negative cases.
• Figure 5: Visualization of attention-weighted concept discovery using CLEAR-HPV.(A) For representative HPV-positive and HPV-negative WSIs from TCGA-HNSCC, we show, in four columns: (i) the original H&E whole slide image, (ii) the $h$-space spatial concept map, (iii) the high-attention spatial concept map, and (iv) regions of interest (ROIs) with their corresponding concept-fraction distributions produced by our CLEAR-HPV. (B) Representative tiles for two CLEAR-HPV concepts (C5 and C7), along with the average concept-fraction vectors for HPV-positive and HPV-negative slides of the entire dataset. We use different colors to represent different concepts consistently across all figures. Blue markings visible on the whole slide images in (A) are pre-existing annotation artifacts on physical slides by clinicians and can be omitted from interpretation.