Confidence-Calibrated Dual-Branch Detection of Oral Cancer from Tongue and Lips Images

Abstract

In order to detect oral cancer early from photos of the tongue and lips, this research introduces a confidence-calibrated, dual-branch framework. A lightweight texture branch (MLBP/HOG) maintains micro-texture, a global CNN encodes colour-shape context, and an attention gate fuses branches per image. Since pixel-level annotations are unavailable, we guide the model’s attention using CAM-consistency regularization to improve lesion localization under weakly supervised training. Improved cross-site robustness is achieved through domain-adversarial alignment, while probability outputs are calibrated through temperature scaling. With stratified evaluation, the model achieves the following on the Oral Cancer (Lips & Tongue) dataset: Brier 0.092, Accuracy 0.892, Macro-F1 0.883, AUROC 0.912, AUPRC 0.884, and ECE reduces from 0.067 to 0.031 after calibration. Low post-calibration ECE (0.029/0.033) and high site-wise performance (Lips AUROC 0.922; Tongue 0.902) are maintained. By combining the texture branch, CAM-consistency, and domain alignment, ablation demonstrates cumulative benefits: when compared to a baseline CNN, the combined performance is the best with minimal compute overhead (AUROC 0.872; AUPRC 0.834; ECE 0.050). When considering utility, a threshold θ* = 0.50 equals Includes a PPV of 0.846, NPV of 0.897, Coverage of 87.2%, and Referral of 12.8%; Sensitivity of 0.892; and Specificity of 0.852. Trustworthy triage is supported by the system's calibrated probabilities and CAM overlays, and real-world deployment on cloud or mobile platforms is encouraged by its robustness to site variability. Practical and reliable photo-based oral-cancer screening relies on complementary features, targeted regularization, and explicit calibration, according to the results.