💻 computer science

CLEAR-HPV: Interpretable Concept Discovery for HPV-Associated Morphology in Whole-Slide Histologyhttps://arxiv.org/submit/7596892/preview

The paper introduces CLEAR-HPV, an interpretable framework that restructures attention-based multiple instance learning to automatically discover and quantify key morphologic concepts (keratinizing, basaloid, and stromal) in HPV-associated whole-slide histology, thereby reducing high-dimensional features to compact, interpretable vectors while maintaining predictive accuracy across diverse cancer datasets.

Original authors: Weiyi Qin, Yingci Liu-Swetz, Shiwei Tan, Hao Wang

Published 2026-05-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Weiyi Qin, Yingci Liu-Swetz, Shiwei Tan, Hao Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a massive, high-resolution photograph of a city (a Whole Slide Image of a tissue sample). This photo is so detailed it contains millions of tiny bricks (cells). Your goal is to figure out if this city was built by a specific type of construction crew (HPV-positive) or a different one (HPV-negative).

Currently, computers are very good at looking at this whole city and saying, "Yes, this is Crew A," or "No, this is Crew B." But they are like black boxes: they give you the answer, but they can't explain why. They just say, "Trust me, I know." Doctors need to know which bricks or neighborhoods led to that conclusion to trust the computer.

This paper introduces a new tool called CLEAR-HPV. Think of it as a translator that takes the computer's "black box" answer and turns it into a simple, understandable story.

Here is how it works, using simple analogies:

1. The Problem: The "Black Box" City

The computer looks at the millions of tiny bricks (tiles) in the photo. It uses a special system called Attention to decide which bricks are important. It's like a security guard who points a flashlight at the most suspicious-looking bricks.

The Issue: The computer knows where to shine the light, but it doesn't have a name for what it sees. It just sees "important brick #452" and "important brick #99." It can't tell a human, "Oh, that's a 'basaloid' brick," or "That's a 'keratinizing' brick."

2. The Solution: The "Concept" Translator

CLEAR-HPV steps in after the computer has already made its decision. It doesn't try to be smarter than the computer; it just tries to organize what the computer already knows.

The Analogy: Imagine the computer has a giant, messy pile of 1,536 different colored marbles (features) for every city. It's impossible for a human to understand a pile of 1,536 colors.
The Magic: CLEAR-HPV looks at the marbles the computer thought were most important (the ones the flashlight shone on). It groups them into just 10 distinct "Concepts" (like 10 specific buckets).
- Bucket 1 might be "Basaloid bricks" (typical of HPV-positive).
- Bucket 2 might be "Keratinizing bricks" (typical of HPV-negative).
- Bucket 3 might be "Inflammatory cells."

3. The Output: A Simple Report Card

Instead of showing the doctor a messy pile of 1,536 numbers, CLEAR-HPV gives them a Concept-Fraction Vector.

The Analogy: This is like a recipe card. Instead of listing every single grain of sand in the city, it says: "This city is made of 40% Basaloid bricks, 10% Keratinizing bricks, and 50% other stuff."
This is powerful because it reduces a complex, confusing 1,536-dimensional problem down to just 10 easy-to-understand numbers.

4. What They Found (The Results)

The researchers tested this on three different groups of patients (datasets) from different hospitals and with different staining methods.

Consistency: Even though the "cities" looked different (different stains, different scanners), the 10 Concepts stayed the same. The "Basaloid" bucket always looked like Basaloid, and the "Keratinizing" bucket always looked like Keratinizing. This means the computer isn't just memorizing the specific photos; it's learning real biological patterns.
Accuracy: When they used these simple 10-number recipes to predict the HPV status, they got almost the exact same accuracy as the original complex computer model. They didn't lose any "brainpower" by simplifying the explanation.
Survival: They also found that these simple recipes could predict how long patients might survive, even though the computer was only trained to guess HPV status. It seems the "Basaloid" vs. "Keratinizing" mix holds clues about the patient's future health.

5. Why This Matters

The paper claims that CLEAR-HPV is the first tool to automatically find these meaningful "Concepts" without needing a human pathologist to label every single brick beforehand.

Before: The computer says, "It's HPV Positive." (Doctor: "Why?") Computer: "I don't know, I just calculated it."
After: The computer says, "It's HPV Positive." (Doctor: "Why?") Computer: "Because 60% of the important bricks in this city are 'Basaloid' style, which is a known sign of HPV."

Summary

CLEAR-HPV is a "post-hoc" (after-the-fact) translator. It takes a powerful but confusing AI model, listens to what it is paying attention to, and organizes that attention into 10 clear, named categories (like "Basaloid" or "Keratinizing"). It proves that you can have a computer that is both smart (accurate) and honest (explainable) without needing to retrain the whole system or hire a human to label millions of images first.

Technical Summary: CLEAR-HPV

Problem Statement

Human papillomavirus (HPV) status is a critical determinant for prognosis and treatment in head and neck and cervical cancers. While HPV-positive tumors typically exhibit non-keratinizing or basaloid morphology and HPV-negative tumors display keratinizing squamous morphology, reliable inference by human observers is difficult due to substantial morphologic overlap. Current computational pathology methods, particularly attention-based Multiple Instance Learning (MIL), achieve strong slide-level prediction but function as "black boxes." They provide limited interpretability, offering only coarse attention heatmaps that indicate where a model attends but failing to explain what specific histologic patterns (concepts) drive the prediction. This lack of concept-level, human-understandable explanation hinders clinical adoption and biological insight.

Methodology: CLEAR-HPV

The authors introduce CLEAR-HPV (Concept-Level Explainable Attention-guided Representation for HPV), a post-hoc framework designed to restructure the latent space of trained attention-based MIL models to enable annotation-free concept discovery.

Core Mechanism

Latent Space Restructuring: The framework operates on the intermediate "h-space" (tile-level embeddings) generated by a pre-trained MIL backbone (e.g., CLAM) before slide-level aggregation.
Attention-Guided Clustering: Instead of clustering raw encoder features, CLEAR-HPV utilizes the attention weights ( $\alpha_i$ ) learned by the MIL model to weight the tile embeddings ( $h_i$ ). This emphasizes diagnostically informative tiles while suppressing background variation.
Concept Discovery: The framework applies unsupervised clustering (specifically k-means variants) to these attention-weighted embeddings to discover $K$ $K$ discrete morphologic concepts (e.g., $K=10$ $K = 10$ ).
- CLEAR-HPV (AW-h): Uses an attention-weighted k-means objective where high-attention tiles exert stronger influence on concept boundaries.
- CLEAR-HPV (raw-h): Clusters the raw h-space embeddings without attention weighting.
Outputs:
- Concept-Fraction Vectors: Each slide is represented by a compact, $K$ -dimensional vector quantifying the proportion of tiles assigned to each concept.
- Spatial Concept Maps: Visualizations showing the spatial distribution of discovered concepts across the Whole Slide Image (WSI).
- Representative Tiles: High-attention tiles closest to concept centroids, providing visual evidence for each concept.

Evaluation Protocol

The framework was evaluated across three independent cohorts: TCGA-HNSCC, TCGA-CESC, and CPTAC-HNSCC. The evaluation focused on:

Predictive Retention: A rule-based "concept-fraction classifier" was trained on the discovered concept vectors to predict HPV status. The goal was to demonstrate that these low-dimensional, interpretable vectors could recover the predictive performance of the original high-dimensional (1536-dim) MIL embeddings.
Zero-Shot Transfer: Models trained on TCGA-HNSCC were applied to external cohorts (TCGA-CESC, CPTAC-HNSCC) without retraining to assess generalization.
Prognostic Utility: The concept vectors were tested for their ability to predict patient survival outcomes, despite being trained solely on HPV labels.

Key Contributions

First Annotation-Free Concept Discovery for HPV: CLEAR-HPV is the first general framework to automatically discover pathology-relevant morphologic concepts for HPV prediction without requiring tile-level annotations or human concept labels.
Attention-Guided Latent Reorganization: The work demonstrates that the latent h-space of attention-based MIL models contains rich, biologically structured information that can be reorganized into discrete, interpretable concepts by leveraging the model's own attention weights.
Dimensionality Reduction with Signal Preservation: The framework reduces high-dimensional feature spaces (e.g., 1536 dimensions) to a compact set of ~10 interpretable concepts while preserving the discriminative signal required for accurate slide-level prediction.
Architectural and Cohort Robustness: The method is compatible with diverse attention-based MIL backbones (CLAM, ABMIL, TransMIL, MHMIL) and demonstrates stable concept discovery and predictive retention across diverse datasets with varying staining protocols and anatomical sites.

Results

Predictive Performance: On TCGA-HNSCC, CLEAR-HPV variants achieved performance comparable to the original CLAM backbone (e.g., AUC ~0.84–0.88 vs. 0.86 for CLAM) using only 10 concept dimensions.
Concept Coherence: The discovered concepts aligned with established biological patterns. Specifically, Concept 5 (C5) consistently corresponded to basaloid/non-keratinizing morphology (enriched in HPV-positive cases), while Concept 7 (C7) corresponded to keratinizing morphology (enriched in HPV-negative cases).
Cross-Cohort Generalization: In zero-shot transfers to TCGA-CESC and CPTAC-HNSCC, the concepts maintained their morphologic identity. While ranking performance (AUC) showed modest degradation under extreme domain shifts, the models preserved high precision and the ability to distinguish morphologic patterns.
Prognostic Signal: The concept-fraction vectors, learned solely for HPV prediction, retained prognostic information for patient survival in TCGA-HNSCC, achieving competitive AUC (~~0.78) compared to the base model (~~0.84).
Baseline Comparisons: CLEAR-HPV (AW-h) and (raw-h) significantly outperformed baselines such as heatmap-based grouping, Dirichlet concept models, and encoder-feature clustering in terms of class-coherent organization and predictive recovery scores.

Significance and Claims

The paper claims that CLEAR-HPV provides a "general, backbone-agnostic framework" that bridges the gap between high-performance deep learning and clinical interpretability. By restructuring the attention-mediated latent space, the authors demonstrate that:

Interpretability is achievable without sacrificing accuracy: The framework offers concept-level explanations (identifying specific morphologic patterns like basaloid vs. keratinizing) while maintaining the predictive power of the original black-box model.
Morphologic structure is inherent in MIL: The latent spaces of attention-based MIL models already encode biologically meaningful organization; CLEAR-HPV simply makes this structure explicit.
Stability across domains: The discovered concepts are not dataset-specific artifacts but reflect stable, transferable biological signals associated with HPV status, generalizing across different cancer types (head/neck vs. cervical) and technical settings.

The authors position CLEAR-HPV not as a method to improve classification accuracy, but as a tool to expose the internal reasoning of MIL models, thereby fostering trust and enabling biological insight in computational pathology.