University of Twente Student Theses


Deep learning segmentation of 3D ultrasound imaging of the thyroid

Munsterman, Roxane (2023) Deep learning segmentation of 3D ultrasound imaging of the thyroid.

[img] PDF
Abstract:Thyroid nodules have a high prevalence and can be detected upon ultrasound imaging. To improve the diagnosis and treatment of thyroid nodules, a 3D segmentation method for ultrasound scans was developed, segmenting the thyroid, carotid artery (CA), and jugular vein (JV). The goal of the method is to aid needle-based interventions, such as radiofrequency ablation (RFA) and improve volumetry accuracy. A tracked sweep dataset from an online repository was used together with a dataset acquired with matrix transducer, which allows for fast 3D volume acquisition. Both datasets consisted of ultrasound scans and annotations from 27 subjects. Pre-processing techniques were applied to enhance the scans, including voxel size normalization and speckle reduction. A U-Net was trained with different strategies (2D, 2.5D majority vote, and 3D) on both the matrix dataset and tracked sweep dataset, to find the best training strategy. The Dice similarity coefficient (DSC) and Hausdorff Distance 95% (HD95) were used to assess the model’s performance. The volume of the prediction was compared to the ground truth and to volumes obtained using the ellipsoid formula. The results showed variations in performance among the training strategies. The 2D model achieved the best results for the tracked sweep dataset in terms of median DSC (0.934, 0.924, 0.897) and HD95 (1.206, 0.588, 1.571 mm) for the thyroid, CA and JV respectively. For the matrix dataset, the 3D train strategy gave overall 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. The model demonstrated lower median volume errors (4.45%) compared to the ellipsoid formula (13.84%) for thyroid volume estimation in the tracked sweep dataset. For the matrix dataset, an error of 7.40% was achieved. A graphical user interface was developed for visualization and clinical use of the segmentation results. A 3D segmentation method for ultrasound volumes of the thyroid, CA and JV was developed. This work paves the way for the development of a planning and navigation method to be used with RFA for thyroid nodules.
Item Type:Essay (Master)
Faculty:TNW: Science and Technology
Subject:44 medicine
Programme:Biomedical Engineering MSc (66226)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page