AAFNet: An oral potentially malignant disorder image segmentation network based on adjacent feature fusion

Oral potentially malignant disorders (OPMD) represents the early stages of precancerous lesions. Accurate diagnosis of OPMDs is crucial for early detection and intervention of oral cancer. However, the clinical diagnosis of OPMDs relies heavily on doctors’ experience and has a certain misdiagnosis rate. In order to relieve doctors’ labor and increase the detection efficiency of OPMDs, automatic approaches to OPMD image segmentation with computer vision and deep learning techniques have been developed. Nonetheless, existing approaches cannot meet the clinical need due to subtle tissue differences and significant lesion variations. So we continue to deal with OPMD image segmentation task. To address these limitations and precisely segment the abnormal tissue regions in OPMD images, we construct an oral lesion dataset OCID containing 1826 images (1277 for training, 183 for validation, and 366 for testing) with normal, benign and OPMD images of various types, shapes and sizes of lesions, accurately annotated by professional doctors, and propose a novel attention mechanism enhanced network AAFNet that includes an adjacent joining module to fuse multi-scale feature maps from adjacent layers with spatial and channel attention mechanisms and a merge guided module to emphasize the target region. Experiments demonstrate that AAFNet achieves a mean IoU of 66.55%, significantly outperforming other state-of-the-art networks. This implies that AAFNet could serve as a valuable computer-aided diagnostic tool for clinicians, and has the potential adapted to other oral lesion types for wider accessibility. This work also lays a solid foundation for future research, including its adaptation to other oral lesion types for wider accessibility. Codes are available at: https://github.com/dec1st2023/AAFNet .