Histological triage of early-stage mycosis fungoides using a weakly supervised deep learning-based model: a multicentre, external validation, and clinical utility study
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Histological diagnosis of early-stage mycosis fungoides (MF) is hindered by profound overlap with benign inflammatory dermatoses (BIDs), leading to diagnostic delays and extensive ancillary testing. We developed MIMIC (Multiple Instance-learning for Identification of Mycosis fungoides In Cutaneous biopsies), a weakly supervised deep learning model designed as a triage tool at initial HCE whole-slide image (WSI) review to distinguish classic patch- and plaque-stage MF from BIDs. We externally validated the model and evaluated its clinical utility.
Methods
In this retrospective multicentre study, we trained a base model using weakly supervised attention-based multiple-instance learning on 3,339 WSIs from two Dutch centres. Crucially, all MF training labels were derived from a deeply phenotyped national cohort featuring strict multidisciplinary expert panel consensus diagnoses (the clinical gold standard). Transportability was evaluated on 371 WSIs from four independent European centres. A blinded reader study on 171 WSIs compared morphology-only performance of MIMIC with 11 (dermato-)pathologists. We then retrained an updated model on all retrospective multicentre data and assessed clinical utility in a strictly held-out, consecutive Utrecht cohort (2022-2023; 486 accessions, 863 WSIs). Primary analysis focused on classic MF versus BIDs (453 accessions). Decision curve analysis, using Platt-scaled probabilities to correct for spectrum bias, evaluated net benefit at a prespecified, safety-oriented threshold of 0.04.
Findings
The base model showed good multicentre transportability (mean centre-specific AUROC 0.91; pooled AUROC 0.84). In the reader study, MIMIC achieved an AUROC of 0.87, exceeding the mean pathologist AUROC (0.79) and the best individual reader (0.83). In the consecutive MF versus BID cohort, the updated model achieved an AUROC of 0.87 (95% CI 0.81-0.92). At the 0.04 threshold, sensitivity was 97.8% (44/45 MF cases) and specificity 50.2%, reducing unnecessary ancillary workups by 39.9 per 100 screening cases versus a test-all strategy.
Interpretation
By identifying nearly half of BIDs as low risk while preserving near-complete sensitivity for classic early-stage MF in a European digital pathology workflow, this unimodal HCE approach offers a scalable digital solution to reduce defensive ancillary testing and accelerate the diagnostic journey for patients with MF. Further validation is needed in non-European centres and in populations with darker skin phototypes.
Funding
This work was supported by an unrestricted Fellowship grant of Stichting Hanarth Fonds in The Netherlands.
Research in context
Evidence before this study
Mycosis fungoides (MF) is notoriously difficult to diagnose in its early stages due to significant clinical and histopathological overlap with benign inflammatory dermatoses (BIDs), resulting in a median diagnostic delay of 36 months from symptom onset. Standard diagnostic workups are highly resource-intensive and frequently inconclusive, often requiring multiple consecutive biopsies, extensive immunohistochemistry panels, molecular T-cell receptor clonality studies, and iterative clinicopathological correlation in multidisciplinary consensus meetings.
We systematically searched PubMed from database inception up to May 28, 2026, using the terms (“mycosis fungoides” OR “cutaneous T-cell lymphoma” OR “CTCL”) AND (“deep learning” OR “machine learning” OR “artificial intelligence” OR “AI” OR “foundation model”) with no language restrictions, identifying 123 raw citations. Only eight original research articles were found to explicitly evaluate machine learning or artificial intelligence applications for cutaneous lymphomas. This pool was supplemented with international conference proceedings and a preprint, identifying a total of thirteen relevant studies. The vast majority of this limited literature focused on non-histological modalities (e.g., clinical photography, dermoscopic imaging, 3D skin tumour burden scoring, or structured electronic health records) or required specialized, non-standard imaging hardware such as non-linear optical microscopy. Furthermore, alternative computational pipelines relied on highly resource-intensive and costly molecular data, such as bulk RNA-sequencing gene-expression classifiers.
Crucially, fewer than five peer-reviewed studies focused on standard brightfield haematoxylin and eosin (HCE) whole-slide images (WSIs) to differentiate early-stage MF from BIDs. While recent baseline architectures, exploratory conference abstracts, and emerging multimodal systems integrating histopathology with clinical metadata have established the existence of a trainable discriminative signal, severe translational limitations persist. Most existing frameworks relied on small development datasets, lacked rigorous external geographic validation across independent international registries, and failed to strictly isolate patient-level clustered data during performance evaluation. Crucially, no prior study has evaluated a unimodal brightfield HCE pathology foundation model within a formalized, Platt-scaled calibration framework to correct for spectrum bias, or utilized decision curve analysis to demonstrate safe, high-sensitivity clinical triage in a consecutive screening pipeline.
Added value of this study
We developed, externally validated, and evaluated the clinical utility of MIMIC, a clinical decision support tool utilizing a pathology foundation model (H-Optimus-1) for risk-stratified triage at the point of initial histological review. In an external geographic validation across four independent European institutions, the base model demonstrated high transportability with a mean centre-specific AUROC of 0·91. In a blinded multi-centre reader study on 171 WSIs, MIMIC achieved an AUROC of 0·87, outperforming the mean baseline performance of 11 independent (dermato-)pathologists (AUROC 0·79 ± 0·03) and exceeding the integrated discriminative area of the top-performing human expert (AUROC 0·83). Crucially, following TRIPOD+AI guidelines for model updating, the final candidate was evaluated on a strictly held-out, consecutive clinical screening workflow cohort ( N = 453 accessions). Following Platt scaling to robustly calibrate for spectrum bias, MIMIC achieved a diagnostic sensitivity of 97·8% at a conservative, safety-oriented operational threshold of 0·04, intercepting 44 out of 45 true malignant cases while effectively identifying 50·2% of BIDs as low-risk on initial HCE morphology alone.
Implications of all the available evidence
MIMIC can be seamlessly integrated into digital pathology workflows to safely “de-bulk” the diagnostic workload by reliably identifying low-risk cases that can be triaged away from automatic ancillary testing. Our automated three-tier workflow analysis demonstrates that a staggering 45·5% of the total diagnostic screening volume can be safely classified as low-risk triage zones where ancillary testing could be avoided, translating to substantial direct laboratory cost savings. Decision curve analysis demonstrated a consistent positive net benefit at the 0·04 threshold, corresponding to a net reduction of 39·9 unnecessary ancillary workups and expert multidisciplinary reviews per 100 screening cases relative to a standard “test all” clinical strategy. By acting as a highly sensitive, objective, and reproducible filter at the earliest stage of histopathological review, this unimodal HCE approach offers a scalable, zero-marginal-cost digital solution to mitigate defensive diagnostic reflexes, optimize specialized laboratory resource allocation, and significantly accelerate the diagnostic pathway for patients with early-stage MF.
