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

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 learning strategy and integrate the denoiser with a segmentation model to produce the task-oriented noiseless estimation. Experimental results show that the proposed method can improve both the visual quality and the segmentation accuracy on spine images.

1. Introduction

In clinical scoliosis diagnosis, experts need to view hundreds of frames in an ultrasound sequence of a whole spine column, which is time-consuming and tedious [1]. To simplify the measurement, Volume Projection Imaging (VPI) was proposed to synthesize coronal 2D images based on the intensity of the voxels in the ultrasound sequence [2, 3]. However, owing to the fast movement of the probe and the noise in the collected spatial information, ultrasound VPI images usually suffer from a significant degradation by structured noise, which not only affects the performance of automatic pathological analysis, but also poses challenges to doctors for accurate diagnosis. As presented in Fig. 1, the structured noise, different from random noise, shows high spatial correlation, and only appears in some regions in the image. The existence of structured noise degrades the discriminative patterns in the ultrasound images, and consequently, confuses the deep network when performing classification, detection, or segmentation. VPI image restoration is an open problem, where the ground-truth data is generally inaccessible. Moreover, the structured noise varies with the ultrasound operators, probes, and empirical imaging parameters, which makes the degradation hard to model. Hence, it is also impractical to synthesize the paired noisy and noiseless samples for learning.