Synergistic Multi-Magnification Fusion Network for Tongue Image-Based Oral Cancer Diagnosis

Oral cancer represents a critical global public health concern, where accurate and timely early detection is paramount. While deep learning on non-invasive tongue and lip images shows potential, single-magnification models fail to capture both macro-level context and micro-level features. Existing multi-magnification fusion methods often suffer from feature misalignment, high computational cost, and poor noise robustness. To overcome these limitations, we propose a novel, computationally efficient deep learning framework for robust multi-magnification fusion in oral lesion screening. Our framework synergistically integrates two lightweight components: 1) a KL-divergence-based feature constraint and 2) a parameter-free max-feature aggregation operator. The KL constraint explicitly enforces distributional consistency across features extracted at different magnifications, effectively mitigating the semantic conflicts. Simultaneously, max aggregation functions as a highly noise-resilient operator, preserving the strongest discriminative activations (which often correspond to critical lesion signals) while suppressing background noise. We rigorously evaluated our framework against other single-magnification and multi-magnification fusion meth ods across three clinically diverse, multi-device tongue and lip image datasets. Our approach is the State-of-the Art (SOTA) strategy in automated oral cancer screening. This performance breakthrough confirms the clinical effectiveness of our synergistic multi-magnification network based on KL constraint-based semantic alignment and max aggregation-based feature selection. Subsequently, we also analyzed and demonstrated the effectiveness of this method in terms of feature discriminability, learning efficiency, and model robustness. Our code is available at https://github.com/DiaoMLab/MM_Fusion-OC.