To explore the clinical application value of deep learning based ultrasonic thyroid nodule segmentation.

We selected 1044 ultrasound images of 166 thyroid patients collected from Beijing Tiantan Hospital Affiliated to Capital Medical University from August 2018 to October 2020. The segmentation effect of Unet with improved self-attention mechanism and control Unet-based method was assessed using test datasets. Whether the segmentation result is close to the manual annotation by a sonographer with many years of clinical experience was used as a reference standard, and the Unet and Unet-based methods with improved self-attention mechanism were compared for the segmentation effect of thyroid nodules, using IoU (intersection and union ratio), Dice (Dice similarity coefficient), and the degree of closeness to the manual outline of thyroid nodules by the sonographer to evaluate the performance and clinical value of the deep learning model for thyroid nodule segmentation.

The IoU and Dice coefficients of thyroid nodule segmentation by Unet with improved self-attention mechanism were 0.815 and 0.839, respectively, which were higher than those of Unet (IoU=0.788, Dice=0.817). It can also be seen from the segmented images that the Unet based on the improved self-attention mechanism had a better segmentation effect on the overall and edge details of thyroid nodules than the Unet-based method, and was closer to the manual outline results of the sonographer.

Unet based on self-attention mechanism has good performance in thyroid nodule segmentation, which can improve the diagnostic efficiency, the method also has clinical application value.