Attention-Gated Feature Fusion in Lightweight Mobilenetv3-Efficientformerv2 for Oral Squamous Cell Carcinoma Classification at High Magnification

Abstract

The diagnosis of Oral Squamous Cell Carcinoma (OSCC) relies heavily on high-magnification (400×) histopathology image analysis to evaluate microscopic cellular details. However, deep learning models frequently suffer from “spatial myopia” at this resolution, where the narrow field of view causes Convolutional Neural Networks (CNNs) to overfit on local textural noise and Vision Transformers (ViTs) to lose global tissue context. To overcome these limitations, this paper proposes an attention-gated feature fusion strategy integrated into a lightweight dual-stream hybrid architecture combining MobileNetV3 and EfficientFormerV2. The proposed Attention Gating module dynamically computes channel-wise probability weights, adaptively filtering out irrelevant global noise while preserving crucial local cellular structures. Evaluated exclusively on a high-magnification $\mathbf{4 0 0} \times$ OSCC dataset utilizing a rigorous 5-fold cross-validation scheme, the proposed model achieves a superior mean diagnostic accuracy of 87.79% and an F1-Score of 85.50%. This comprehensively outperforms its standalone constituent backbones and modern state-of-theart lightweight baselines, including EfficientFormerV2 (F1: 83.32%) and EfficientViT-M2 (F1: 81.62%). Requiring only 17.53 million parameters, the results demonstrate that the attention-gated mechanism effectively mitigates spatial myopia, providing a highly robust, interpretable, and computationally efficient Computer-Aided Diagnosis (CAD) solution suitable for clinical edge-computing deployments.