CLARA-OSCC: A Calibrated Lesion-Aware Resolution- Adaptive Pipeline for Oral Squamous Cell Carcinoma Diagnosis

Abstract

Background: Early diagnosis of Oral Squamous Cell Carcinoma (OSCC) from histopathological images is challenging because of limited and imbalanced datasets, variations in image magnification (100× and 400×), and calibration degradation associated with site-specific staining and scanner shifts. This study proposes CLARA-OSCC, a magnification-aware framework designed to improve robust classification, calibration, and clinically relevant triage of OSCC under heterogeneous imaging conditions.

Materials and Methods: CLARA-OSCC integrates robust image standardization, a multi-resolution texture-spectral lattice, a resolution gate, a lightweight SE-MobileViT backbone, gated cross-modal feature fusion, class-balanced focal learning, and domain alignment using CORAL. Post-hoc temperature scaling is employed to improve prediction calibration. The framework is evaluated using internally stratified sub-samples of a publicly available OSCC histopathological image dataset, with separate assessment at 100× and 400× magnifications. Performance is assessed using Accuracy, F1-score, AUROC, AUPRC, Brier score, Expected Calibration Error (ECE), Sensitivity, Specificity, PPV, and NPV.

Results: CLARA-OSCC achieved an Accuracy of 0.958 ± 0.024, F1-score of 0.959 ± 0.026, AUROC of 0.985 ± 0.009, and AUPRC of 0.978 ± 0.012 at the 95% confidence interval. The framework obtained a Brier score of 0.033 and ECE of 0.021, representing improvements of 0.011 and 0.015, respectively, over the baseline. At the selected clinical threshold of θ = 0.80, the framework achieved 85% coverage, with Sensitivity of 0.951, Specificity of 0.969, PPV of 0.971, and NPV of 0.948. Performance was higher at 400× magnification, with an F1-score of 0.966 and ECE of 0.018, compared with 0.949 and 0.027, respectively, at 100× magnification. Reliability analysis demonstrated reduced overconfidence following calibration, while prototype-based retrieval provided case-level explanations.

Conclusion: The results indicate that magnification-aware feature fusion combined with explicit calibration can provide robust and triage-oriented predictions for OSCC histopathological images. The framework also provides mechanisms for handling simulated domain variation, uncertainty, and abstention-based routing. Although the site-level analysis used internally stratified sub-samples of a publicly available dataset, the observed robustness provides motivation for future multicentre external validation before clinical deployment.