Deep learning-based segmentation of 3D ultrasound images of the thyroid

The goal of the study was to develop a method for segmentation of the thyroid, carotid artery (CA), and jugular vein (JV) using 3D ultrasound data. This method forms the basis for a computer-assisted needle-based intervention for thyroid nodules and thyroid volume estimation accuracy. Two datasets were used: the first was acquired using a tracked 2D sweep and the second with a 3D matrix transducer. A 2D and 3D U-Net model were trained on the full data set with different strategies (2D, majority vote in 2.5D and 3D). The 2D model achieved the best results for the tracked 2D sweep data set in terms of median Dice Score Coefficient (DSC) (0.934, 0.924, 0.897) and Hausdorff distance at the 95 percentile (HD₉₅) (1.206, 0.588, 1.571 mm) for the thyroid, CA, and JV, respectively. For the matrix data set, the 3D model gave overall the best results in its median DSC (0.869, 0.930, 0.856) and HD₉₅ (1.814, 0.606, 1.405 mm) for the thyroid, CA, and JV, respectively, showing comparable results in vessel segmentation but inferior results in thyroid segmentation compared to the tracked sweep data set. The model demonstrated lower median volume estimation errors in the tracked sweep data set (4.45 %) compared to the matrix data set (7.40 %) and the ellipsoid formula (13.84 %) for thyroid volume estimation. This work shows that automatic segmentation in 3D ultrasound of the human neck is best performed with 3D ultrasound. Improving the quality of the 3D data is important for the development of a planning and navigation method to be used with needle-based interventions for thyroid nodules.