MiFDeU: Multi-information fusion network based on dual-encoder for pelvic bones segmentation

Abstract

The segmentation of bone fragments is crucial for preoperative planning and intraoperative navigation in reduction surgery. Recent advances in medical segmentation have predominantly focused on U-shaped frameworks that employ convolutional neural networks or transformer variants as the backbone. However, these frameworks, which rely on a single encoder, often struggle with integrating information from diverse features and processing irregular shapes in visual objects. Such limitations can reduce segmentation accuracy and impair generalization performance across different datasets. To address these issues, we introduce a multi-information fusion network based on dual-encoder for pelvic bones segmentation. In order to capture global contextual information and local features simultaneously, our model takes a light resnet and a graph neural network with swin-pool module as dual-encoder for effectively representing the global and local topologies. We construct a high-low multi-dimensional paired attention in the bottleneck for fusing spatial and channel information from different dimensions. Instead of using the traditional dice loss in the unet-like architecture, our model employs both topological loss and boundary loss to enhance the goal optimization. In the experiments, our model achieves a substantially lower dice similarity coefficient and comparable 95 % Hausdorff distance compared to other state-of-the-art. The experiments on across datasets verify the superiority and generalization of the proposed model.