Author: Zilei Fu, Yi Guo, Wanli Huo, Hongdong Liu, Laishui Lyu, Zhao Peng, Yaping Qi, Senting Wang 👨🔬
Affiliation: Department of Radiotherapy, cancer center, The First Affiliated Hospital of Fujian Medical University, the Zhejiang-New Zealand Joint Vision-Based Intelligent Metrology Laboratory, College of Information Engineering, China Jiliang University, Division of lonizing Radiation Metrology, National Institute of Metrology, State Key Laboratory of Oncology in South China, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangdong Provincial Clinical Research Center for Cancer, Sun Yat-sen University Cancer Center, China Jiliang University, Department of Oncology, Xiangya Hospital, Central South University 🌍
Purpose: Medical image boundaries are commonly characterized by smooth gray-level transitions, resulting in pixel-level segmentation errors near these blurred boundaries. To address this, we developed the MGA-UNet model, which uses a unique MGA module to capture both global coarse-grained and fine-grained features, thereby enhancing the accuracy of multi-organ segmentation.
Methods: We have designed and implemented the MGA-UNet model. Specifically, the Multi-Level Large Kernel Attention (MLKA) in MGA can extract features at different granularity levels, capturing both global coarse-grained and fine-grained features. Combined with the Gated Spatial Attention Unit (GSAU) and feature fusion module, it maintains the ability to model global context. The Multi-Scale Convolutional Attention (MSCA) module connects and enhances feature maps extracted at different scales through convolutions, establishing dependencies between multi-level features. The multi-grained features extracted by MLKA and GSAU are aggregated with the multi-scale features extracted by MSCA. Limitations in imaging device resolution, patient movement, and artifacts during imaging can result in pixels near boundaries being prone to segmentation errors, often leading to incomplete boundaries and label overflow in segmentation results. The aforementioned method can reduce errors caused by CT imaging artifacts while perceiving the overall organ shape.
Results: The MGA-UNet model demonstrated exceptional segmentation performance on Synapse and ACDC datasets, tackling complex segmentation tasks with high-precision boundary segmentation. It accurately delineated organs like the liver and kidneys and distinguished adjacent, structurally similar tissues. Avg. DSC up 4.51% on Synapse, 0.69% on ACDC vs. classics.
Conclusion: This work presents an innovative segmentation algorithm based on multi-granularity feature extraction technology, showing remarkable performance in medical image segmentation tasks. Future research will deepen the application of multi-granularity feature extraction in medical image segmentation and explore its performance under different conditions, aiming to further optimize algorithm performance and improve segmentation accuracy and reliability.