Uncertainty-aware MSI triage from H&E slides works only where confidence intervals hold

Why MSI triage matters and where AI fits in the diagnostic pathway

Microsatellite instability (MSI) is a key predictive biomarker in colorectal cancer, present in approximately 15% of cases, and its accurate determination guides immunotherapy selection and Lynch syndrome screening [2]. Reference-standard testing relies on PCR, IHC, or NGS, which show high concordance (Kappa 0.930–0.986) but remain resource-intensive and unevenly available [4]. Universal MSI/MMR testing is endorsed by major international societies, yet the workload and cost of molecular testing create a bottleneck that AI-based pre-screening from routine H&E slides could potentially alleviate [8]. Earlier deep learning studies demonstrated that MSI status is predictable from H&E morphology, with a median AUC consistently above 0.90 across 82 identified studies, but these models typically output point estimates without uncertainty quantification or abstention mechanisms [8].

The clinical logic of triage differs fundamentally from diagnosis. In a triage scenario, the model must identify MSS cases with high confidence so they can safely skip or delay molecular testing, while all other cases—including uncertain ones—proceed to standard testing [1]. This safety-oriented design prioritizes negative predictive value over overall accuracy, because misclassifying an MSI case as MSS could deny a patient immunotherapy [1]. The new paper explicitly adopts this framing, reporting coverage (the proportion of cases receiving a definitive MSS call), sensitivity (the proportion of MSI cases not misclassified as MSS), and NPV (the probability that an MSS prediction is correct) as the primary metrics [1].

How attention-based MIL, ensembling, and confidence intervals work together

The anchor model builds on attention-based multiple instance learning (ABMIL), a framework in which a whole-slide image is divided into patches, each patch is encoded by a foundation model, and an attention mechanism aggregates patch-level features into a slide-level prediction [1]. Earlier work by Schirris et al. demonstrated that contrastive self-supervised pre-training combined with heterogeneity-aware MIL improved MSI prediction from 0.77 to 0.87 AUROC on TCGA-CRC, establishing that both representation learning and aggregation strategy matter [7]. The new paper introduces three architectural modifications: using register tokens instead of the class token from the penultimate block of the H-Optimus-0 foundation model, restricting classification to the top 10 attended patches per ensemble member, and training an ensemble of 100 models whose predictions are averaged [1].

The ensemble serves two purposes. First, it improves predictive performance: the combined AUC across all external cohorts was 0.965 (CI 0.955–0.974) for the proposed PB-Reg-Top10 model, compared with 0.937 (CI 0.924–0.950) for the baseline Last-CLS-All configuration [1]. Second, and more importantly for triage, the ensemble enables estimation of epistemic uncertainty through bootstrapped confidence intervals (CIens) [1]. A prediction is classified as 'Undetermined' if the mean ensemble score falls between decision thresholds or if the CIens overlaps either threshold, and these cases are forwarded to conventional molecular testing [1]. This abstention mechanism is what transforms a probabilistic classifier into a clinically deployable triage tool, and it directly addresses the governance of indeterminate predictions identified as an unresolved challenge in systematic reviews [8].

External cohort performance: high sensitivity but variable coverage

Across five independent external cohorts (ATB, NIB, ING, PAIP, TR), the model achieved sensitivities above 97% for MSI detection, but coverage—the proportion of cases classified as MSS—ranged from 59% to 78% [1]. In the ATB cohort, 359 of 440 MSS cases were correctly classified, with only 5 MSI cases misclassified as MSS, yielding 59% coverage and 97.7% sensitivity [1]. The NIB cohort showed similar coverage (59%) with 98.0% sensitivity, while ING, PAIP, and TR achieved 78%, 60%, and 78% coverage respectively, with 100% sensitivity in all three [1]. Negative predictive values were consistently high, ranging from 98.0% to 100% across cohorts [1].

These results mean that in the best-performing cohorts, approximately one in five cases could potentially avoid immediate molecular testing, while in the more challenging cohorts, only about three in five cases receive a definitive MSS call. The variability in coverage across cohorts—despite consistent sensitivity—suggests that cohort-specific factors such as slide quality, tumor purity, or preprocessing differences affect the model's confidence calibration [1]. The authors note that MSI prevalence in their evaluation cohorts (12–40%) exceeded the ~15% typical in consecutive CRC, which was a deliberate choice to establish reliable sensitivity estimates but may affect NPV and coverage under real-world sampling [1].

Comparison with competing approaches and the limits of multimodal prediction

The anchor paper's performance compares favorably with the publicly released Wagner et al. model, which achieved a combined AUC of 0.940 (CI 0.926–0.954) versus 0.965 for the proposed model, despite being trained on approximately 11,000 cases versus 1,242 [1]. The Wagner model performed slightly better on the NIB cohort (AUC 0.979 vs. 0.974) but markedly worse on ATB (0.910 vs. 0.959) and PAIP (0.921 vs. 0.993), suggesting that the ensemble and uncertainty-aware approach generalizes more consistently across institutions [1]. However, the authors acknowledge that incomplete reproducibility of the original preprocessing pipeline complicates direct comparison [1].

Competing multimodal approaches have reported higher AUCs by incorporating additional data sources. Tang et al. developed a fusion model combining contrast-enhanced CT and histopathology, achieving an external test AUC of 0.905 [3]. Ao et al. reported a multiomics nomogram integrating CT radiomics, pathomics, and clinical variables with AUCs of 0.996–0.999 in training and internal validation, and 0.993 in external validation [5]. These higher numbers, however, come with important caveats: the Ao et al. model requires preoperative CT imaging and clinical variables, which may not be available in all settings, and the extremely high AUCs raise questions about overfitting or cohort-specific calibration that the anchor paper's uncertainty framework is designed to detect [5]. The anchor paper's contribution is not superior discrimination but rather a framework for knowing when the model's prediction should be trusted [1].

Boundaries: what the model cannot do and what remains uncertain

The most important limitation is that the model cannot replace molecular testing. The high-sensitivity triage covers only 40–76% of cases, meaning that 24–60% of cases are classified as Undetermined and require standard PCR or IHC [1]. The authors explicitly state that high-confidence MSI prediction may be valuable in research but should not be relied upon for clinical decision-making without further prospective validation [1]. Additionally, the study was restricted to resection specimens; application to biopsies—which are clinically important for neoadjuvant immunotherapy decisions—remains unvalidated and presents specific challenges including smaller tumor area and fragmented tissue [1].

Several open questions remain. First, the quality control step relies on accurate tissue segmentation to estimate tumor area but does not assess overall specimen quality, and manual quality assurance as performed in this study may not be feasible in clinical practice [1]. Second, MSI prevalence in the evaluation cohorts exceeded typical consecutive CRC prevalence, so NPV and coverage may differ under real-world sampling [1]. Third, evaluation was conducted at the slide level; patient-level aggregation strategies were not explored and may affect clinical workflows when multiple slides per patient are available [1]. Fourth, molecular annotation beyond MSI/MMR status was available only for the NIB cohort, limiting characterization of discordant cases across cohorts [1]. Finally, while the approach does not require cohort-specific calibration, performance could potentially be improved through calibration if sufficient data is available [1]. The broader challenge of uncertainty quantification in medical imaging—balancing computational cost, calibration quality, and clinical interpretability—remains an active area of research [6].

About These Sources

This research page is built on 8 peer-reviewed studies — published from 2022 to 2026, 7 from 2024 or later — selected as the most relevant from 13 studies that passed quality screening, drawn from 100 papers retrieved from a database of over 500 million.

Sources used in this answer