MAFR-UNet: multi-scale adaptive feature reassembly network for aortic CTA segmentation

The combinations of Convolutional Neural Networks (CNNs) and Transformer have shown promising results in many medical image segmentation tasks. However, the simple feature stacking or concatenation may neglect multi-scale semantic alignment, which induces semantic ambiguity during the decoding phase. Moreover, in many existing CNN and Transformer hybrid U-shaped segmentation models, the skip connections directly transfer features within corresponding network layers, which may introduce unexpected noise and undermine the global representation advantage of Transformer. In this paper, we introduce MAFR-UNet, a model that effectively integrates CNN and Transformer to synergize their local and global feature extraction capabilities. Specifically, the backbone is enhanced by a Multi-scale Adaptive Feature Reassembly (MAFR) module to capture and align multi-level features. Meanwhile, a residual CNN bottleneck module is incorporated to strengthen local feature extraction while reducing computational complexity. To address the ambiguity of target boundaries in CTA images, we incorporate a specific boundary-aware loss term into the objective function, thereby explicitly enforcing sharper boundary delineation. Experimental results demonstrate that the MAFR-UNet achieves competitive segmentation performance. Furthermore, the model exhibits good generalization capabilities across multi-organ segmentation tasks, highlighting its potential for diverse clinical applications. The source code and implementation details of MAFR-UNet are available at https://github.com/xnwn/MAFR-UNet.