DCNN models with post-hoc interpretability for the automated detection of glossitis and OSCC on the tongue

Abstract

This study aimed to develop and evaluate deep convolutional neural network (DCNN) models with Grad-CAM visualization for the automated classification with interpretability of tongue conditions—specifically glossitis and oral squamous cell carcinoma (OSCC)—using clinical tongue photographs, with a focus on their potential for early detection and telemedicine-based diagnostics. A total of 652 tongue images were categorized into normal control (n = 294), glossitis (n = 340), and OSCC (n = 17). Four pretrained DCNN architectures (VGG16, VGG19, ResNet50, ResNet152) were fine-tuned using transfer learning. Model interpretability was enhanced via Grad-CAM and sparsity analysis. Diagnostic performance was assessed using AUROC, with subgroup analysis by age, sex, and image segmentation strategy. For glossitis classification, VGG16 (AUROC = 0.8428, 95% CI 0.7757–0.9100) and VGG19 (AUROC = 0.8639, 95% CI 0.7988–0.9170) performed strongly, while the ensemble of VGG16 and VGG19 achieved the best result (AUROC = 0.8731, 95% CI 0.8072–0.9298). OSCC detection showed near-perfect performance across all models, with VGG19 and ResNet152 achieving AUROC = 1.0000 and VGG16 reaching AUROC = 0.9902 (95% CI 0.9707–1.0000). Diagnostic performance did not differ significantly by age (P = 0.3052) or sex (P = 0.4531), and whole-image classification outperformed patch-wise segmentation (P = 0.7440). DCNN models with Grad-CAM demonstrated robust performance in classifying glossitis and OSCC from tongue photographs with interpretability. The results highlight the potential of AI-driven tongue diagnosis as a valuable tool for remote healthcare, promoting early detection and expanding access to oral health services.