LAUSR

Autonomous ultrasound scanning robots have attracted the attention of researchers, and the real-time quality assessment of ultrasound images is the key technology of them. Existing robot systems usually use pixel-level feature statistical methods such as grayscale, confidence map, etc. However, in clinical practice doctors’ evaluation of ultrasound image quality not only relies on the pixel quality, but also on the image content. In this study, we introduced the deep learning method to the quality assessment of medical breast ultrasound images to learn the doctors’ clinical evaluation standards. We collected 1205 breast ultrasound images of 533 patients and asked experienced doctors to score them. The ResNet18 with a shallow number of layers is adopted to extract features of ultrasound images, and the high-order feature coding model BCNN (Bilinear CNN) in fine-grained categorization is adopted to improve the accuracy of quality assessment. The consistency between the ultrasound image quality assessment results of our method and the doctors’ annotation results reached 0.842 by PLCC (Pearson Linear Correlation Coefficient). Compared with the confidence map method commonly used for quality assessment in the automatic ultrasound scanning robots, our method achieves a higher consistency with the evaluation results of doctors and provides a new image evaluation method and idea for the visual servo control framework of ultrasound autonomous scanning robots.