Research and Application of Tongue Image Segmentation Method Based on Lightweight Deep Learning

Intelligent tongue diagnosis demands high-precision segmentation on resource-constrained mobile devices. All deep learning models outperform conventional approaches, improving mean intersection over union (mIoU) by approximately 9% and accelerating inference by more than two orders of magnitude. Among the evaluated models, UNet++ achieves the highest test mIoU (0.9543) and Dice (0.9764), with paired bootstrap analysis confirming statistically significant advantages over BiSeNet, PP-LiteSeg, and HRNet. BiSeNet achieves the fastest inference speed (310.03 FPS) and the highest mPA (0.9766), making it the most efficient model for real-time deployment. HRNet and PP-LiteSeg provide a favorable accuracy–efficiency balance, suitable for mobile-based tongue diagnosis. Deployment of PP-LiteSeg on three Android devices achieved inference latency of 233–262 ms and throughput of 3.2–3.4 FPS, demonstrating the feasibility of lightweight models on commodity mobile hardware. These findings establish an evidence-based framework for model selection in mobile tongue diagnosis, supporting AI deployment on portable devices for clinical assistance and remote monitoring.