Lightweight AI Models for Robust Oral Cancer Screening

Oral cancer poses a significant health challenge, particularly in low-to-middle-income countries where specialist shortages delay diagnosis. Smartphone-based screening offers a scalable solution, but developing robust AI for such resource-constrained environments faces hurdles like data imbalance, variable quality, and computational limits on edge devices. This research focuses on optimizing lightweight deep learning models for this critical application. The study utilized a large, diverse multi-center dataset of approximately 30,000 images collected over a decade. Researchers systematically evaluated state-of-the-art convolutional, transformer, and hybrid architectures. Through rigorous pipeline ablation, they found that directly optimizing hybrid architectures for edge deployment significantly outperformed computationally heavier approaches, such as using large models or knowledge distillation. The optimized MobileViTv2 models achieved impressive results on a held-out test set, with an average sensitivity of 83.2% and specificity of 86.0%. The best model reached 87.4% sensitivity and 86.5% specificity, alongside a critical negative predictive value of 97.2% against specialist labels. Interpretability analysis confirmed that the system relies on relevant clinical features, and simulated noise tests showed robustness to unstructured sensor noise, despite some vulnerability to impulse bit errors. These findings confirm the high potential of interpretable and robust lightweight AI for automated triage in primary care.