Skip to main content

Weakly supervised deep learning predicts perineural invasion from H&E whole-slide images in tongue squamous cell carcinoma: development and external validation

In brief

AI model predicts nerve invasion in tongue cancer with 0.94 external AUC

In a retrospective study of 629 patients, a model analyzing routine tissue slides predicted perineural invasion with an area under the curve of 0.94 in an external test cohort of 192 patients. It could help flag slides for focused pathologist review, but prospective studies must show whether it improves workflow or postoperative decisions.

Journal
International journal of medical informatics (Q1)
Published
3 November 2026
Study design
Prospective / inception cohort
Evidence level
Level 2, Moderate (CEBM 2b)
Authors
Zhe Shen, Yike Zhang, Jianze Yu, Hang Ling, Pingqing Tan, Liang Gong
PMID
42998450
DOI
10.1016/j.ijmedinf.2026.106781

Why clinicians should know about it

Abstract

BACKGROUND: Perineural invasion (PNI) in tongue squamous cell carcinoma (TSCC) is associated with aggressive behavior and poor outcomes, yet it can be under-recognized on H&E sections because it is often focal and interpretation varies. Manual assessment of PNI suffers from significant interobserver variability, motivating AI-based decision support. We developed and externally validated a weakly supervised WSI model to predict PNI in TSCC. METHODS: We retrospectively analyzed 1,511 digitized H&E whole-slide images from 629 patients at two centers. Patient-level partitioning yielded an internal training cohort of 305 patients and 787 WSIs, an internal validation cohort of 132 patients and 340 WSIs, and an external test cohort of 192 patients and 384 WSIs. Non-overlapping tiling generated 41,552,500 candidate patches before tissue-background exclusion, of which 7,813,555 unique tissue patches were retained for analysis. A two-stage pipeline was used: patch-level modeling and MIL-based slide aggregation. ResNet50 was selected over Inception_v3 and SimpleViT as the feature extractor. Patch-level prediction probabilities and image features were aggregated into a 101-dimensional patch likelihood histogram (PLH) and a 101-dimensional visual bag-of-words (BoW) representation, respectively. For conventional classifiers, the two representations were concatenated into a 202-dimensional vector, followed by z-score normalization and LASSO feature selection. For the Transformer, the two representations were processed separately using PCA, independently projected into 128-dimensional embedding spaces, and integrated using four-head cross-attention. Performance was assessed by ROC, decision curve analysis (DCA), and calibration. RESULTS: The internal training, internal validation, and external test cohorts included 305, 132, and 192 patients, respectively. PNI prevalence was 82.3%, 79.5%, and 72.9% in these cohorts, respectively. The Transformer fusion model achieved AUCs of 0.990 in the training cohort, 0.968 in the internal validation cohort, and 0.940 in the external test cohort. Calibration and decision curve analysis were used to characterize the model's predicted probabilities and modeled net benefit in the retrospective study cohorts. Patch-level Grad-CAM was used to generate qualitative saliency visualisations associated with ResNet50 patch-level PNI scores. CONCLUSIONS: This externally validated computational pathology application demonstrated retrospective technical feasibility for WSI-level PNI prediction in TSCC and has potential clinical utility for flagging slides for focused pathologist review. Its effects on clinical workflow and postoperative decision-making require prospective evaluation.

Abstract as published, via PubMed.

View on PubMedFull text at the publisherOpen in the app

For healthcare professionals. The summary is generated by AI from the published abstract, and the evidence level is assigned automatically from the study design on the Oxford CEBM hierarchy. Neither is medical advice. Read the full paper before changing practice.