An artificial intelligence model of whole-slide pathology specimens differentiating cutaneous high-grade squamous proliferations
Authors:
- Anne PetzoldAPAnne Petzold
- Anja WesselyAWAnja Wessely
- Michael ErdmannMichael Erdmann
- Stefan SchliepSSStefan Schliep
- Stephan SchremlSSStephan Schreml
- Luis Carlos Rivera MonroyLMLuis Carlos Rivera Monroy
- Julio VeraJVJulio Vera
- Konstantin DrexlerKDKonstantin Drexler
- Dennis NiebelDNDennis Niebel
- Kinan Maurice HayaniKHKinan Maurice Hayani
- Franklin KiesewetterFKFranklin Kiesewetter
- Carola BerkingCBCarola Berking
Open Access
- 25 September 2025
- journal article
- research article
- Published by Springer Nature in Virchows Archiv
- Vol. 487 (5) , 1047-1058
Abstract
Cutaneous squamous cell carcinoma (cSCC) and verruca vulgaris (VV) are skin conditions involving the proliferation of epidermal keratinocytes requiring fundamentally different treatments. Histological evaluation of highly differentiated squamous cell proliferations can be challenging, particularly in small or superficial samples. This study aims to improve diagnostic accuracy using an AI model to distinguish cSCC from VV. We developed a deep-learning model using clustering-constrained attention multiple instance learning (CLAM) to classify hematoxylin and eosin-stained whole-slide images (WSIs) as cSCC or VV. The dataset comprised 289 WSIs (n = 148 cSCC, n = 141 VV). Quality control was ensured through expert review: the training cohort was evaluated by four dermatopathologists, and the evaluation cohort by six additional experts. On the training set, the model achieved an AUROC of 0.99, with an accuracy of 94.9% for cSCC and 91.2% for VV. On the evaluation set, it reached an AUROC of 0.96, and accuracies of 82.4% (cSCC) and 97.4% (VV), similar to the average performance of individual dermatopathologists. We successfully trained and implemented an interpretable deep-learning-based weakly supervised model on WSIs distinguishing cSCC from VV, which could enhance AI-supported diagnostics in the future.Keywords
Funding Information
- Universitätsklinikum Erlangen
This publication has 27 references indexed in Scilit:
- Artificial Intelligence in Dermatology: A Systematic Review of Its Applications in Melanoma and Keratinocyte Carcinoma DiagnosisDermatologic Surgery, 2024
- Deep learning based histological classification of adnex tumorsEuropean Journal of Cancer, 2023
- European consensus-based interdisciplinary guideline for invasive cutaneous squamous cell carcinoma: Part 2. Treatment–Update 2023European Journal of Cancer, 2023
- Scaling Self-Supervised Learning for Histopathology with Masked Image ModelingmedRxiv, 2023
- Recommendations on compiling test datasets for evaluating artificial intelligence solutions in pathologyLaboratory Investigation, 2022
- An attention-based weakly supervised framework for spitzoid melanocytic lesion diagnosis in whole slide imagesArtificial Intelligence in Medicine, 2021
- Data-efficient and weakly supervised computational pathology on whole-slide imagesNature Biomedical Engineering, 2021
- Deep learning in digital pathology image analysis: a surveyFrontiers of Medicine, 2020
- Man against machine: diagnostic performance of a deep learning convolutional neural network for dermoscopic melanoma recognition in comparison to 58 dermatologistsAnnals of Oncology, 2018
- Dermatologist-level classification of skin cancer with deep neural networksNature, 2017