DA-GAN: Learning Structured Noise Removal In Ultrasound Volume Projection Imaging For Enhanced Spine Segmentation

Abstract

Ultrasound volume projection imaging (VPI) has shown to be appealing from a clinical perspective, because of its harmlessness, flexibility, and efficiency in scoliosis assessment. However, the limitations in hardware devices degrade the resultant image content with strong structured noise. Owing to the unavailability of reference data and the unpredictable degradation model, VPI image recovery is a challenging problem. In this paper, we propose a novel framework to learn the structured noise removal from unpaired samples. We introduce the attention mechanism into the generative adversarial network to enhance the learning by focusing on the salient corrupted patterns. We also present a dual adversarial strategy and integrate the denoiser with a segmentation model to produce the task-oriented noiseless estimation. Experimental results show that the proposed method can greatly improve both the visual quality and the segmentation accuracy on spine images.

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

“…We consider the reason to be that the strong scan noise in the VPI images corrupts the discriminative patterns of spine bones, which restricts those methods from fully investigating the discriminative features for segmentation. Our proposed framework also surpasses those previous studies especially designed for spine segmentation, i.e., DAGAN [16] and SEAM…”

“…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. However, it performs restoration and segmentation independently.…”

“…Banerjee et al [26] proposed a lightweight UNet to perform effective spine segmentation with a low computational burden. In our previous study, we proposed a generative adversarial network with dual adversarial learning (DAGAN) to perform noise removal, which is cascaded with a segmentation network for enhanced spine segmentation [16]. However, DAGAN suffers from low efficiency because the segmentation model has to wait for the denoised image produced by the noise removal network.…”

“…This paper is an extension of our early work in [16]. We improve our previous method by designing a selective featuresharing multi-task framework to replace the original cascaded learning strategy.…”

A dual channel and spatial attention network for automatic spine segmentation of MRI images

“…Cheng et al 14 developed a two‐stage Dense‐U‐Net based on a deep learning method for automatic positioning and segmentation of the vertebral body to diagnose lumbar spinal stenosis (LSS). Huang et al 25 proposed a deep‐learning framework, which removed noise from unpaired samples. ResAttenGAN 26 achieves better segmentation by adding a residual refinement attention module to refine the segmentation boundary.…”