医学图像分割中的双分支特征提取器及高效特征融合方法

In medical image segmentation, convolutional and Transformer networks are highly favored for their unique advantages, but their respective applications also face specific limitations. In addition, existing feature fusion modules suffer from significant information loss and cannot fully learn and utilize the complex relationships between space and channels to achieve more accurate segmentation. To this end, a dual branch parallel network feature extractor is first proposed, which solves the shortcomings of a single network in information extraction and effectively overcomes the information bottleneck problem that may occur when two networks are combined in series. Meanwhile, to fully utilize the complex relationship between space and channels, this paper further introduces a multi-branch local global feature fusion enhancement module, which can efficiently fuse features from both branches. The experiments show that the algorithm performs well on the Synapse and ACDC datasets, with an average Dice of 83.32% and 91.82%, and an HD95 index of 15.80 mm and 1.29 mm, respectively, demonstrating strong competitiveness.