Towards Cross-Scale Attention and Surface Supervision for Fractured Bone Segmentation in CT

TL;DR

The paper tackles automated fracture segmentation in CT by addressing boundary delineation and large morphological variability. It introduces a cross-scale attention (CSA) mechanism to fuse multi-scale features and a surface supervision loss that emphasizes accurate bone boundaries within a 3D U-Net framework. On a hip-fracture pelvic CT dataset, the approach achieves notable performance ($DSC=93.36\%$, $ASSD=0.85$ mm, $95HD=7.51$ mm) and outperforms graph-cut, 3D U-Net, and Transformer baselines, with ablations confirming the value of both CSA and surface supervision. The method's boundary-focused improvement and strong results suggest it can generalize to other fracture types and be combined with different segmentation backbones.

Abstract

Bone segmentation is an essential step for the preoperative planning of fracture trauma surgery. The automated segmentation of fractured bone from computed tomography (CT) scans remains challenging, due to the large differences of fractures in position and morphology, and also the inherent anatomical characteristics of different bone structures. To alleviate these issues, we propose a cross-scale attention mechanism as well as a surface supervision strategy for fractured bone segmentation in CT. Specifically, a cross-scale attention mechanism is introduced to effectively aggregate the features among different scales to provide more powerful fracture representation. Moreover, a surface supervision strategy is employed, which explicitly constrains the network to pay more attention to the bone boundary. The efficacy of the proposed method is evaluated on a public dataset containing CT scans with hip fractures. The evaluation metrics are Dice similarity coefficient (DSC), average symmetric surface distance (ASSD), and Hausdorff distance (95HD). The proposed method achieves an average DSC of 93.36%, ASSD of 0.85mm, 95HD of 7.51mm. Our method offers an effective fracture segmentation approach for the pelvic CT examinations, and has the potential to be used for improving the segmentation performance of other types of fractures.

Figures (5)

• Figure 1: Illustration of bone CT slices from three subjects. It can be observed that due to the inherent anatomical characteristics of the bone structures (the adjacent region of multi-bones in (a) and the narrow and thin regions in (b)), and the various types of fractures in (c), the automated segmentation remains challenging.
• Figure 2: The network architecture of the fracture segmentation model. The backbone is based on 3D U-net work14. The cross-scale attention (CSA) modules are used to enhance the feature representation. Conv: convolution; BN: batch normalization; ReLU: rectified linear unit.
• Figure 3: Detailed structure of the proposed cross-scale attention (CSA) module. The W, H, D, C denote the width, height, depth and channel number of the feature map, respectively. The $Conv$ is the convolution operation with the kernel size $k_{size}$.
• Figure 4: 3D visualization of the segmentation results from different methods.
• Figure 5: 2D visualization of the segmentation results on hip fractures.