Unsupervised Ultrasound Image Quality Assessment with Score Consistency and Relativity Co-learning

Selecting an optimal standard plane in prenatal ultrasound is crucial for improving the accuracy of AI-assisted diagnosis. Existing approaches, typically dependent on detecting the presence of anatomical structures as defined by clinical protocols, have been constrained by a lack of consideration for image perceptual quality. Although supervised training with manually labeled quality scores seems feasible, the subjective nature and unclear definition of these scores make such learning error-prone and manual labeling excessively time-consuming. In this paper, we present an unsupervised ultrasound image quality assessment method with score consistency and relativity co-learning (CRL-UIQA). Our approach generates pseudo-labels by calculating feature distribution distances between ultrasound images and high-quality standard planes, leveraging consistency and relativity for training regression networks in quality prediction. Extensive experiments on the dataset demonstrate the impressive performance of the proposed CRL-UIQA, showcasing excellent generalization across diverse plane images.