🤖 machine learning

RACR-MIL: Rank-aware contextual reasoning for weakly supervised grading of squamous cell carcinoma using whole slide images

The paper proposes RACR-MIL, a weakly-supervised attention-based multiple-instance learning framework that utilizes a hybrid WSI graph and rank-ordering constraints to achieve state-of-the-art generalization and accuracy in grading squamous cell carcinoma across multiple anatomical sites.

Original authors: Anirudh Choudhary, Mosbah Aouad, Krishnakant Saboo, Angelina Hwang, Jacob Kechter, Blake Bordeaux, Puneet Bhullar, David DiCaudo, Steven Nelson, Nneka Comfere, Emma Johnson, Olayemi Sokumbi, Jason Slu

Published 2026-08-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Anirudh Choudhary, Mosbah Aouad, Krishnakant Saboo, Angelina Hwang, Jacob Kechter, Blake Bordeaux, Puneet Bhullar, David DiCaudo, Steven Nelson, Nneka Comfere, Emma Johnson, Olayemi Sokumbi, Jason Sluzevich, Leah Swanson, Dennis Murphree, Aaron Mangold, Ravishankar Iyer

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

Cancer diagnosis often begins with a pathologist peering through a microscope at a thin slice of tissue, searching for subtle clues that reveal how aggressive a tumor might be. For a common type of skin and organ cancer called squamous cell carcinoma, this task is notoriously difficult. The disease does not look the same everywhere; a single sample can contain areas that are slow-growing and others that are rapidly spreading, all mixed together. Traditionally, doctors assign a grade to the entire sample based on the worst-looking area they can find, a process that relies heavily on human experience and can vary from one doctor to another. To help standardize this, scientists have turned to artificial intelligence, teaching computers to read whole-slide images of tissue. However, teaching a computer to distinguish between different levels of severity is tricky because the computer is usually shown only the final diagnosis for the whole slide, without being told exactly which tiny spots within that slide are the most dangerous.

A team of researchers has developed a new approach to solve this problem, creating a system that mimics the way a human expert thinks when grading these tumors. Instead of treating every tiny piece of the tissue image as an isolated fact, their method, called RACR-MIL, builds a map of how different parts of the tissue relate to one another. It understands that the behavior of a cancer cell depends on its neighbors and its broader environment. More importantly, the system is designed to prioritize the most dangerous areas. Just as a human pathologist knows that a small patch of highly aggressive cells can define the severity of the entire case, this AI is trained to look for and weigh those specific high-risk regions more heavily than the calmer, less dangerous areas. By combining a deep understanding of tissue context with a strict rule to focus on the worst parts, the system achieves a level of accuracy that surpasses previous methods.

The researchers tested this new framework on thousands of tissue samples from patients with squamous cell carcinoma in the skin, lungs, and head and neck regions. They compared their system against a wide range of existing artificial intelligence tools, including those that simply look at individual patches of tissue or those that try to connect them in more basic ways. The results were clear: the new system consistently outperformed its competitors. It improved the accuracy of grading by a significant margin, correctly identifying the severity of the cancer in cases where other methods struggled. Perhaps most importantly, it was better at finding the exact locations of the most dangerous tumor cells. In tests, it located these critical regions up to ten percent more effectively than the next best methods, a difference that could be vital for a doctor trying to decide on a treatment plan.

What makes this approach unique is how it handles the complexity of the tissue. The system constructs two different types of connections between the tiny image patches. One type links patches that are physically close to each other, capturing the immediate neighborhood where cells interact, such as how a tumor pushes against surrounding healthy tissue. The other type connects patches that look similar, even if they are far apart on the slide, allowing the system to recognize patterns of cancer growth that might be scattered across the sample. The researchers found that using both types of connections together provided a much richer picture than using either one alone. Furthermore, the system includes a specific mechanism that forces it to pay more attention to patches that show signs of severe disease. This ensures that the final diagnosis is driven by the most critical evidence, rather than getting lost in the average appearance of the whole slide.

To see if this technology could actually help in a real clinic, the researchers invited two expert pathologists to review a set of cases with and without the help of the AI. In the majority of cases, the doctors found that the system helped them work more efficiently. It successfully highlighted the most aggressive parts of the tumors that the doctors had initially missed or underestimated, prompting them to revise their diagnoses to a more accurate, higher grade. In one instance, the system identified a small, highly dangerous area within a larger, less aggressive tumor that the doctor had overlooked. While there were a few cases where the system's focus on the worst areas made it miss some nuances of the surrounding tissue, the overall feedback was positive. The doctors felt the tool acted as a reliable assistant, pointing them toward the most important parts of the slide and reducing the time needed to reach a confident conclusion.

The study also looked at how well the system would work on data it had never seen before, such as tissue samples from different hospitals or scanned with different machines. The new framework proved to be remarkably robust, maintaining its high performance even when the data changed. This suggests that the system has learned the fundamental rules of how these cancers behave, rather than just memorizing the specific images it was trained on. The researchers noted that while the system is not perfect and still requires human oversight, it represents a significant step forward in making cancer grading more consistent and reliable. By aligning the computer's logic with the way human experts prioritize information, this work offers a promising path toward better, faster, and more accurate cancer diagnoses for patients around the world.

Technical Summary: RACR-MIL for Weakly Supervised SCC Grading

1. Problem Statement

Squamous cell carcinoma (SCC) is a pervasive cancer subtype with significant mortality impact, arising in diverse anatomical sites including the skin, head and neck, and lung. Accurate histological grading of SCC from whole slide images (WSIs) is critical for prognosis and clinical management but remains challenging due to:

Tumor Heterogeneity: Significant intratumoral variation and subtle morphological differences between adjacent grades (well, moderate, poor).
Lack of Standardization: The absence of a reliable, standardized grading protocol.
Contextual Dependencies: Grading relies on integrating cellular differentiation with broader tissue context (e.g., keratinization, stromal reaction, tumor growth patterns) across multiple scales.
Weak Supervision Limitations: Existing weakly supervised learning (WSL) methods often fail to explicitly model long- and short-range tissue dependencies or consistently prioritize diagnostically critical, higher-grade tumor regions. Standard attention-based Multiple Instance Learning (MIL) can dilute attention across varying grades, treating slides with mixed grades equivalently and failing to localize the most aggressive tumor components.

2. Methodology: RACR-MIL

The authors propose RACR-MIL (Rank-aware Contextual Reasoning for MIL), a weakly supervised framework designed to classify WSIs into four classes: normal, well-differentiated, moderately-differentiated, and poorly-differentiated. The architecture integrates a hybrid dual-graph representation with a rank-aware attention mechanism.

2.1. Tissue Feature Encoding

Local Feature Extraction: WSIs are tiled into non-overlapping patches (448×448 at 20X magnification). Features are extracted using self-supervised learning (SSL) via a hierarchical transformer (Nest-S) pretrained with DINO. These features are projected into a lower-dimensional space via an MLP to emphasize grading-specific information.
Hybrid Dual-Graph Construction: To capture contextual dependencies, each WSI is represented as a graph with two complementary adjacency matrices:
1. Latent Graph ( $A_{lat}$ ): Connects morphologically similar patches regardless of spatial distance. It is constructed using k-nearest neighbors (kNN) in feature space, refined via reciprocal kNN to reduce noise, and enhanced using graph diffusion (Personalized PageRank) to capture higher-order relationships.
2. Spatial Graph ( $A_{spa}$ ): Connects physically adjacent patches (8-nearest neighbors) to capture local tumor microenvironment context (e.g., keratinization, immune infiltration).
Contextual Feature Derivation: An attention-based Graph Neural Network (Att-GNN) processes both graphs. Each graph utilizes a distinct self-attention head to learn graph-specific relationships. A diversity regularizer minimizes the correlation between the attention weights of the two graphs, ensuring they learn complementary rather than redundant patterns. The final contextual features are a residual combination of the local features and the aggregated graph features.

2.2. Grade Classification with Rank-Aware Constraints

The model aggregates patch features using attention but introduces specific constraints to mimic the ordinal nature of SCC grading:

Rank-Aware Attention: Instead of standard pooling, the model uses conjunctive pooling with a temperature-scaled cosine classifier.
Inter-Grade Constraint: Enforces that patches associated with higher grades (e.g., poorly differentiated) receive higher attention weights than those associated with lower grades. This is implemented via a pairwise ranking loss on grade-specific attention scores derived from high-confidence patches.
Intra-Grade Constraint: Encourages the model to prioritize patches with higher confidence scores within the same grade class, improving the alignment between patch-level predictions and slide-level aggregation.
Loss Function: The total objective combines the grade classification loss, inter-grade ranking loss, intra-grade ranking loss, and graph diversity loss.

3. Key Contributions

Dual-Graph Representation: The first weakly supervised framework for multiclass SCC grading that jointly models local spatial context and non-local phenotypic relationships through a hybrid dual-graph architecture, enhanced by a diversity constraint.
Rank-Aware Attention Mechanism: Introduction of inter-grade and intra-grade constraints that explicitly prioritize regions exhibiting stronger evidence of higher-grade tumors, aligning the model's feature aggregation with the ordinal diagnostic process of pathologists.
Robust Generalization: Comprehensive evaluation across three distinct SCC cohorts (skin, head & neck, and lung), demonstrating state-of-the-art performance and resilience to class imbalance and domain shifts (out-of-distribution evaluation).
Clinical Validation: A pilot study with pathologists indicating the model's potential as a diagnostic assistant, improving grading efficiency and accuracy in identifying high-risk tumor regions.

4. Experimental Results

The model was evaluated on skin SCC (private dataset), Lung SCC (CPTAC), and Head & Neck SCC (TCGA/CPTAC) using five-fold cross-validation.

Performance Gains: RACR-MIL achieved state-of-the-art performance, outperforming existing MIL baselines (including ABMIL, CLAM-MB, PatchGCN, TransMIL) by 3–9% in grade classification F1 scores and up to 10% in tumor localization sensitivity.
High-Risk Tumor Detection: The model showed significant improvements (up to 28%) in classifying challenging higher-grade (moderate and poor) tumors, which are often minority classes.
Localization: In localization analysis against pathologist annotations, RACR-MIL achieved higher sensitivity and saliency compared to non-contextual (ABMIL) and contextual (PatchGCN) baselines, particularly in correctly identifying poorly differentiated regions.
Ablation Studies:
- The dual-graph formulation consistently outperformed single-graph variants (latent-only or spatial-only).
- Both inter-grade and intra-grade ranking constraints were shown to be complementary; their joint application yielded the best performance.
- The diversity regularizer further improved performance by ensuring the two graph branches captured distinct information.
Feature Backbones: The framework demonstrated robustness across different feature extractors (SCC-DINO, PathGen-CLIP, CONCHv1.5), with consistent gains over baselines.

5. Significance and Claims

The paper claims that RACR-MIL represents a significant step forward in computational pathology for SCC by addressing the specific challenges of tumor heterogeneity and the lack of standardized grading protocols.

Clinical Relevance: The framework aligns with the pathologist's diagnostic process by prioritizing the most aggressive tumor components, a critical factor in SCC grading.
Generalizability: Unlike previous methods often limited to specific cancer types or anatomies, RACR-MIL demonstrates robust generalization across skin, lung, and head/neck SCC, suggesting a unified approach for diverse anatomical sites.
Clinical Utility: In a pilot study, pathologists reported that the model improved grading efficiency in 60% of cases and successfully highlighted missed high-grade regions. The authors position RACR-MIL as a viable AI-assisted tool to support, rather than replace, pathologist workflows.

The authors acknowledge limitations, including the need for statistical significance testing with larger cohorts, potential false positives in non-tumor regions, and the trade-off between patch size and localization granularity. They suggest future work in multi-scale representations, uncertainty quantification, and extension to other cancer types and survival prediction tasks.