Structure-Enhanced Attentive Learning For Spine Segmentation From Ultrasound Volume Projection Images

Abstract

Automatic spine segmentation, based on ultrasound volume projection imaging (VPI), is of great value in clinical applications to diagnose scoliosis in teenagers. In this paper, we propose a novel framework to improve the segmentation accuracy on spine images via structure-enhanced attentive learning. Since the spine bones contain strong prior knowledge of their shapes and positions in ultrasound VPI images, we propose to encode this information into the semantic representations in an attentive manner. We first revisit the self-attention mechanism in representation learning, and then present a strategy to introduce the structural knowledge into the key representation in self-attention. By this means, the network explores both the contextual and structural information in the learned features, and consequently improves the segmentation accuracy. We conduct various experiments to demonstrate that our proposed method achieves promising performance on spine image segmentation, which shows great potential in clinical diagnosis.

Joint Spine Segmentation and Noise Removal From Ultrasound Volume Projection Images With Selective Feature Sharing

“…Apparently, the multi-task algorithm, i.e., DCR [25], the state-of-the-art medical image segmentation method, i.e., nnUNet [54], as well as the spine segmentation method, i.e., DAGAN [16], tend to predict a false alarm of lumbar vertebra at this place. SEAM [24] can produce comparable results with our proposed method to eliminate this false alarm, because SEAM [24] further considers the structure supervision in learning. Our proposed method, benefiting from the proposed selective feature-sharing joint-learning strategy, can accurately predict the spine features without structure supervision.…”

“…We first evaluate our proposed framework on spine segmentation by comparing it with other state-of-the-art segmentation methods under the same settings, including the benchmark methods of UNet [51], FPN [44] and HRNet [52], the state-ofthe-art algorithms of nnUNet [54] and UNet++ [53] for medical image segmentation, the multi-task algorithms of MASSL [55] and DCR [25], and the methods of DAGAN [16] and SEAM [24] especially designed for ultrasound VPI images. It is worth noting that our previous work DAGAN [16] also aims to recover those scan noises in VPI images.…”

“…Precision AHD FPN [44] 0.0001 0.0001 0.0001 nnUNet [54] 0.0001 0.0001 0.0002 DAGAN [16] 0.0001 0.0001 0.0001 [24], by approximately 1.1% and 1.4% respectively on the average Dice score, which shows its superiority in performing spine segmentation from ultrasound VPI images.…”

“…Huang et al [23] proposed an efficient regularization-based algorithm to address the occlusion issue in VPI images for enhanced spine segmentation. Zhao et al [24] proposed to introduce the structure supervision to the representation learning in a self-attention manner for more effective spine segmentation. Lyu et al [25] presented a dual-task framework with boundary detection as an auxiliary task to regularize spine segmentation.…”