Three-dimensional Visualization of Thyroid Ultrasound Images Based on Multi-scale Features Fusion and Hierarchical Attention

Abstract

Background The incidence of thyroid disorders is increasing every year. Three-dimensional visualization of the thyroid gland and its surrounding tissues based on free-arm ultrasound scanning videos is essential for identifying the thyroid and related tissues in the neck and screening for associated diseases. However, the key to achieving 3D visualization lies in the multi-target segmentation of neck tissues. Currently, the accuracy of multi-target segmentation for thyroid ultrasound images are unsatisfactory. Method Therefore, in order to enhance the multi-target segmentation accuracy for the thyroid and its surrounding tissues and improve the 3D visualization, in this paper we proposed PA-Unet++. Pyramid pooling module helps in the segmentation of multi-organization with different sizes by integrating multi-scale feature information. Additionally, attention gating mechanism is applied to each decoding layer to progressively highlight the target tissue and suppress the representation of background pixels. Results With 4600 images containing 23,000 annotated regions are divided into training and test sets at a ratio of 9:1, the results show that: compared with the results of U-net++, the Dice of our model increased from 78.78–81.88% (+ 3.10%), the mIOU increased from 73.44–80.35% (+ 6.91%), the PA index increased from 92.95–94.79% (+ 1.84%). Conclusions This indicates that PA-Unet + + has a higher accuracy for segmentation of multiple tissues in thyroid ultrasound images. The segmentation results can optimize the 3D visualization.