H&E-based MSI/MMR testing with AI in colorectal cancer: a multi-centred blinded evaluation

Cher Bass, Foivos Ntelemis, Julian Schmidt, Steffen Wolf, André Geraldes, Debapriya Mehrotra, Shikha Singhal, Narender Kumar, Angelica Marcia, Nicholas Bennett, Oscar Maiques, Mitchell Hyde, Bejal Mistry, Grace Rogerson, Michele Cummings, Clare Freer, Elizabeth Walsh, Manuel Salto-Tellez, Maurice B. Loughrey, In Hwa Um, David J. Harrison, Richard Clarkson, James Blackwood, J. Carl Barrett, Jakob Nikolas Kather, Nicolas M. Orsi, Pahini Pandya, Salim ArslanNpj Digit. Medicine 2025 · DOI

TL;DR

This paper introduces PANProfiler Colorectal (PPC), an AI-powered tool that analyzes standard H&E-stained tissue slides to determine microsatellite instability (MSI) and mismatch repair (MMR) deficiency in colorectal cancer (CRC). In a large, multi-center blinded study, PPC achieved high accuracy, offering a faster, more accessible alternative to current diagnostic methods.

What Problem Does This Paper Address?

Microsatellite instability (MSI) and mismatch repair (MMR) deficiency are crucial biomarkers in colorectal cancer (CRC), influencing treatment decisions, particularly regarding immunotherapy response and identifying patients for Lynch syndrome testing. Current methods for detecting MSI/MMR status, such as immunohistochemistry (IHC) and polymerase chain reaction (PCR), are effective but have significant drawbacks. These include high costs, long turnaround times, the need for specialized equipment and infrastructure, and potential inter-observer variability in interpretation. These limitations hinder widespread adoption and timely clinical utility, creating a need for a more accessible, cost-effective, and rapid diagnostic solution.

Methods

The core of this research is PANProfiler Colorectal (PPC), an artificial intelligence (AI)-based system designed to predict MSI/MMR status directly from hematoxylin and eosin (H&E)-stained whole slide images (WSIs). PPC employs a deep learning pipeline that involves several key steps:

  • Data Preprocessing: This includes background removal, tissue detection, tiling of the WSI into smaller patches, and color normalization to ensure consistency across different scans.
  • Feature Extraction: A Convolutional Neural Network (CNN) acts as a feature extractor, learning unique representations from each tile.
  • Attention Module: This module aggregates tile embeddings and weights them based on their relevance to the MSI/MMR status, creating a slide-level embedding.
  • Classification: A final classification layer uses the slide-level embedding to predict the biomarker status as 'Unstable' (MSI-high/dMMR), 'Stable' (proficient MMR/non-MSI-high), or 'Indeterminate'.

The study also utilized a self-supervised learning (SSL) framework for pre-training the base model, which was found to outperform a standard ImageNet backbone. Explainability was addressed by visualizing "attention heatmaps" to highlight regions of the WSI that the model focused on for its predictions, and by using a separate tissue classification model to annotate regions.

Datasets

The study utilized several datasets:

  • Development and Calibration:
    • T1-US-CRC-SVS-1-DEV: 563 WSIs from The Cancer Genome Atlas (TCGA).
  • I1-UK-CRC-CZI-1-DEV: 1,156 WSIs from Scotland.
  • L1-UK-CRC-SVS-1-DEV: 551 WSIs from Leeds.
  • N1-UK-CRC-SVS-1-DEV: 130 WSIs from Northern Ireland.
  • Blinded Validation:
    • L1-UK-CRC-SVS-1-BLIND: 3,124 WSIs from England.
  • W1-UK-CRC-CZI-1-BLIND: 54 WSIs from Wales.
  • N1-UK-CRC-SVS-1-BLIND: 398 WSIs from Northern Ireland.

A total of 3,576 WSIs from 1,243 CRC patients were used for the blinded validation.

Key Insights and Ideas

  • H&E is Sufficient: The study demonstrates that standard H&E-stained slides, which are routinely prepared for all cancer biopsies, contain enough morphological information for AI to accurately predict MSI/MMR status. This bypasses the need for additional, specialized testing.
  • AI as a Diagnostic Tool, Not Just Screening: Unlike some previous AI approaches focused on pre-screening, PPC is designed for diagnostic use, aiming for high sensitivity and specificity. This is achieved through a three-tiered classification system ('Stable', 'Unstable', 'Indeterminate') that allows for definitive calls with higher confidence.
  • Explainability Matters: The use of attention heatmaps provides crucial insights into why the AI makes a particular prediction, highlighting specific histological features (e.g., tumor-infiltrating lymphocytes, poor differentiation) that correlate with MSI/MMR status. This builds trust and aids in understanding potential misclassifications.
  • Potential for Workflow Revolution: PPC's ability to provide results in minutes, rather than days, has the potential to significantly streamline the diagnostic workflow, reduce pathologist workload, and enable faster treatment decisions.

Key Results

The blinded validation study yielded strong performance metrics:

  • Overall Agreement (OPA): 93.83%
  • Positive Agreement (PPA): 92.54%
  • Negative Agreement (NPA): 94.02%
  • Test Replacement Rate (TRR): 86.55% of slides yielded a definitive result, meaning only 13.45% would require further standard testing.
  • C-statistic: Consistently high across cohorts, exceeding 0.92.

The performance was robust across different UK institutions, suggesting good generalizability.

Limitations and Future Work

  • Indeterminate Results: Approximately 14% of samples returned an "Indeterminate" result. The authors attribute this to potential preanalytical factors (e.g., suboptimal tissue processing, staining) and morphological inconsistencies.
  • Generalizability: While validated across three UK institutions, further validation with broader geographical and demographic diversity is recommended.
  • Biopsy Samples: The current study primarily used resected tissue samples. Future work will focus on validating PPC with biopsy samples, which are more commonly used in routine clinical practice.
  • Health Economics: A health economics analysis is planned to evaluate the cost-effectiveness of PPC.
  • Usability Studies: User experience and workflow integration will be assessed through usability studies with target users.
  • Standardization: Standardizing preanalytical workflows is crucial to increase the proportion of assessable cases and improve clinical usability.

Summary updated

BibTeX
@article{bass2025h,
  author = {Cher Bass and Foivos Ntelemis and Julian Schmidt and Steffen Wolf and André Geraldes and Debapriya Mehrotra and Shikha Singhal and Narender Kumar and Angelica Marcia and Nicholas Bennett and Oscar Maiques and Mitchell Hyde and Bejal Mistry and Grace Rogerson and Michele Cummings and Clare Freer and Elizabeth Walsh and Manuel Salto-Tellez and Maurice B. Loughrey and In Hwa Um and David J. Harrison and Richard Clarkson and James Blackwood and J. Carl Barrett and Jakob Nikolas Kather and Nicolas M. Orsi and Pahini Pandya and Salim Arslan},
  title = {{H\&E-based MSI/MMR testing with AI in colorectal cancer: a multi-centred blinded evaluation}},
  journal = {Npj Digit. Medicine},
  year = {2025},
  doi = {10.1038/s41746-025-02218-5},
}
Figure