Multi-Dimensional Fusion Reverse Attention Network for Polyp Segmentation

As computer-aided diagnostic technology continues to advance, medical image analysis plays a crucial role in the early detection and treatment of gastrointestinal diseases. Colorectal cancer is a common malignant tumor, and the accurate identification and segmentation of colonic polyps, which are its precursors, are of great significance for the prevention of colorectal cancer. However, due to the diversity in the morphology and color of polyps, as well as the concealment of small or flat polyps, accurately segmenting them in endoscopic images poses a challenge. This paper proposes an innovative Multi-Dimensional Fusion Reverse Attention Network (MdfraNet) aimed at precisely identifying and segmenting the boundaries of polyps and their surrounding mucosa. This method leverages a Multi-Scale Mobile Inverted Convolution Model (MSMB) to achieve deep fusion of multidimensional features, capturing contextual information and detailed features from different scales. Furthermore, the study introduces a Hierarchical Multiscale Cross Reverse Attention (HMC-RA)mechanism, which effectively integrates information from multiple scales and locations, significantly enhancing feature representation, especially in handling polyp boundary information. Extensive testing of MdfraNet on five public polyp datasets (CVC-300, CVC-ColonDB, CVC-ClinicDB, Kvasir-SEG, and ETIS-LaribPolypDB) demonstrates that this model has a significant advantage in handling boundary information compared to existing techniques, showcasing excellent performance. Additionally, MdfraNet has also shown superior segmentation results on our in-house dataset from our collaborative hospital.