Truace
Subscribe

A Statistically Validated and Explainable Deep Learning Framework for Oral Cancer Classification

Despite numerous deep learning models reporting high accuracy for oral cancer classification, most rely on a single train/test split on a single dataset, with limited statistical validation, external generalization testing, or explainability. This study addresses that gap through a statistically validated, externally tested, and dual-explainable classification framework. We provide a rigorous benchmarking study contrasting 20 hybrid combinations (4 pre-trained backbones × 5 classical classifiers) against fine-tu…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-1638· cite & verify
A Statistically Validated and Explainable Deep Learning Framework for Oral Cancer Classification
Clinical Update Vol. 29, No. 3 by U.S. Navy. Naval Postgraduate Dental School. Public domain

In brief

Researchers benchmarked 20 hybrid combinations of four pre-trained backbones and five classical classifiers against fine-tuned end-to-end CNNs for oral cancer classification. Using repeated stratified fivefold cross-validation, ResNet50 with an SVM-RBF classifier performed best, and it was then tested on an independent external dataset.

The work matters because many earlier oral cancer models reported high accuracy from a single split on a single dataset, with little statistical validation or explainability. Adding external testing and visual and feature-level explanations aims to build clinical trust, but the study describes the result only as a foundation toward screening support, not a deployed clinical tool.

Main points

  1. The study compared 20 hybrid backbone-classifier combinations with fine-tuned end-to-end CNNs.
  2. In cross-validation with 10 repetitions, ResNet50 plus SVM-RBF reached 0.960 accuracy, 0.953 F1-score and 0.991 ROC-AUC.
  3. Wilcoxon signed-rank tests confirmed the performance differences were statistically significant (p < 0.001).
  4. On an external dataset, accuracy fell to 0.916, while ROC-AUC was 0.984 and precision was 0.98.
  5. MobileNetV2 with SVM-RBF was identified as a lightweight option for resource-constrained point-of-care settings.

The gain

A hybrid pipeline pairing a pre-trained ResNet50 backbone with an SVM-RBF classifier outperformed end-to-end CNNs on oral cancer classification and kept strong accuracy on an independent external dataset.

The rundown

The authors say most existing oral cancer deep learning models rely on one train/test split on one dataset, with limited statistical validation, external generalization testing or explainability. Their design responds to each of those gaps.

Internal performance was measured with repeated stratified fivefold cross-validation over 10 repetitions, with confidence intervals reported. The top model's ROC-AUC confidence interval was 0.990-0.993, and the hybrid approach beat end-to-end CNNs.

On unseen external samples, the best model's high precision of 0.98 indicates a low false-positive rate. Grad-CAM supplies localized heatmaps and SHAP supplies global feature attribution to support interpretability.

Sources

  1. Peer-reviewedJournal of Imaging Informatics in Medicine2026-10-09

The debate