To explore the feasibility of deep learning-based restoration of obscured thyroid ultrasound images.

A total of 358 images of thyroid nodules were retropectively collected from January 2020 to October 2021 at Sir Run Run Shaw Hospital, Zhejiang University School of Medicine, and the images were randomly masked and restored using DeepFillv2. The difference in grey values between the images before and after restoration was compared, and 6 sonographers (2 chief physicians, 2 attending physicians, 2 residents) were invited to compare the rate of correctness of judgement and detection of image discrepancies. The ultrasound features of thyroid nodules (solid composition, microcalcifications, markedly hypoechoic, ill-defined or irregular margins, or extrathyroidal extensions, vertical orientation and comet-tail artifact) were extracted according to the Chinese Thyroid Imaging Reporting and Data System (C-TIRADS). The consistency of ultrasound features of thyroid nodules before and after restoration were compared.

The mean squared error of the images before and after restoration ranged from 0.274 to 0.522, and there were significant differences in the rate of correctness of judgement and detection of image discrepancies between physicians of different groups(all P<0.001). The overall accuracy rate was 51.95%, the overall detection rate was 1.79%, there were significant differences also within the chief physicians and resident groups (all P<0.001). The agreement rate of all ultrasound features of the nodules before and after image restoration was higher than 70%, over 90% agreement rate for features such as solid composition and comet-tail artifact.

The algorithm can effectively repair obscured thyroid ultrasound images while preserving image features, which is expected to expand the deep learning image database, and promote the development of deep learning in the field of ultrasound images.

Ultrasonography; Thyroid nodules; Image restoration; Deep learning