Multi-View Fusion Network For Malignant Oral Ulcer Lesions Detection

This study aims to develop a multi-view feature fusion network with multi-scale perception capabilities to accurately identify complex and heterogeneous malignant oral ulcer lesions, supporting early detection and diagnosis. A dataset of 800 de-identified clinical photographs containing various oral ulcer lesions was constructed. Images were annotated by clinical experts using specialized image annotation tools to establish a robust training set for deep learning. We proposed a novel multi-view guided feature fusion network incorporating scale-aware mechanisms for precise segmentation and detection of oral ulcer lesions. Specifically, a multi-view embedding module employing parallel multi-path embedding strategies and gated enhancement techniques was introduced to effectively integrate semantic context and detailed edge features. Additionally, a scale-aware multi-attention mechanism combined with hierarchical semantic feature fusion was utilized to optimize the representation of high-level semantic information and low-level fine-grained details in lesion areas, thereby maximizing mutual information utilization and feature extraction capability. The proposed network achieved an average intersection-over-union (mIoU) exceeding 90%, demonstrating superior performance compared to existing state-of-the-art methods in various evaluation metrics, effectively realizing accurate localization and detection of malignant oral ulcers. This research introduces a highly effective deep learning architecture for the detection and segmentation of malignant oral ulcers, significantly enhancing diagnostic accuracy and providing essential support for the early screening and preventive management of oral squamous cell carcinoma.