Multi‐group deformable convolution network for 3D medical image segmentation
Jiasheng Li, Mingzhe Hu, Jing Wang, Tonghe Wang, David S. Yu, Xiaofeng Yang · Medical Physics · 2025
BACKGROUND: Medical image segmentation plays an important role in radiation oncology, including delineating anatomical structures and detecting abnormalities. Precise segmentation of medical images is essential for accurately contouring organs at risk (OAR) and target volumes to ensure effective and safe radiation therapy. PURPOSES: Recently, there has been growing interest in developing Vision Transformer (ViT) or Convolutional Neural Network (CNN) methods for 3D medical image segmentation. These advanced methods are crucial for handling the complex spatial and semantic structure of 3D medical images, which necessitates both large receptive fields and adaptations to varying spatial geometries. However, previous works in both CNNs and ViTs have demonstrated limitations in fully capturing these complexities, highlighting the need for more sophisticated approaches. METHODS: We introduce MGDC-Net, a Multi-Group Deformable Convolution (MGDC) network for 3D volumetric medical image segmentation. Our MGDC-Net employs deformable convolution operators with learnable spatial offsets to improve attention on semantically important regions. Our approach leverages stable spatial distribution across subjects to enhance semantic learning. We also incorporate transformer components to augment feature learning and reduce inductive biases inherent in traditional CNNs. Our network employs a combination of deformable convolution operators and transformer components to improve feature learning and computational efficiency. RESULTS: MGDC-net was evaluated on three diverse segmentation tasks using public datasets: brain tumor segmentation (BraTS21), CT multi-organ segmentation (FLARE21), and cross-modality MR/CT segmentation (AMOS22). MGDC-Net demonstrated superior performance on the three segmentation tasks, achieving 91.4% DSC on brain tumor segmentation (BraTS21), 94.4% DSC on CT multi-organ segmentation (FLARE21), and 84.1% DSC on cross-modality MR/CT segmentation (AMOS22). Our network also compared favourably with existing methods in terms of computational efficiency. CONCLUSIONS: MGDC-Net provides a robust and efficient method for 3D volumetric medical image segmentation, demonstrating significant improvements in segmentation performance across multiple tasks. This network's ability to leverage deformable convolutions and transformer components suggests its potential for broader applications in medical image analysis.