Code for: Evaluating Local-Global Representation in CNN-ViT Hybrid for OSCC Histopathological Classification with Grad-CAM++
Description
Abstract—Oral Squamous Cell Carcinoma (OSCC) is a malignant disease with high mortality that is often diagnosed at an advanced stage. This study systematically evaluates local-global feature representation for OSCC histopathology classification using a CNN-ViT hybrid architecture that integrates EfficientNetB3 with a Transformer Encoder. Rather than focusing solely on architecture comparison, a phased ablation study was conducted to examine the contributions of the Transformer Encoder, data augmentation, and class weighting to classification performance. Grad-CAM++ was further employed as an Explainable AI (XAI) method to examine model attention patterns in histopathological images. Experiments were performed on the Histopathologic Oral Cancer Detection dataset from Kaggle, with multi-seed evaluation using five random seeds. The complete configuration combining the Transformer Encoder, data augmentation, and class weighting achieved the highest observed performance, with 92.86% accuracy, 90.04% F1-score, and 96.84% sensitivity. Across five seeds, the model achieved an average accuracy of 91.27% ± 1.51% and sensitivity of 94.53% ± 1.23%. Grad-CAM++ visualizations showed activation patterns concentrated on histopathological regions relevant to OSCC classification. However, expert pathologist validation is required to establish the clinical relevance of these interpretations.
Files
Institutions
- Muhammadiyah University of YogyakartaYogyakarta, Yogyakarta