上海科技大学知识管理系统

The conventional method to evaluate spinous process angles (SPAs) is to obtain the spinous processes (SP) curve by manually locating spinous processes on radiographs. HR-Net is a deep neural network which uses a multi-feature fusion strategy for keypoints detection. The objectives of this study are to automatically locate the SP on the ultrasound (US) transverse images by applying the High-Resolution network (HR-Net) and then to measure the SPAs on the reconstructed coronal image. The HR-Net model was trained on 1200 US transverse images and tested on 386 images to locate the spinous process. Twenty-five scoliotic subjects were scanned for the evaluation of SPAs measurement. After detecting the SP positions on each frame using HR-Net, the 3D image volumes were reconstructed, and the SPAs were measured on the coronal planes. HR-Net predicted the five keypoints on the test set with the average accuracy of 74.09% with the SP accuracy of 80.05%. The mean absolute difference (MAD) of SPAs between US and radiographic measurement was 2. 7±2.0°, and the correlation was 0.89. The results showed that the HR-Net method could automatically locate the spinous processes on US transverse images and moreover provide accurate estimation of SPAs for scoliotic subjects. © 2021 IEEE.