Optimization of Thyroid Volume Determination by Stitched 3D-Ultrasound Data Sets in Patients with Structural Thyroid Disease

Ultrasound (US) is the most important imaging method for the assessment of structural disorders of the thyroid. A precise volume determination is relevant for therapy planning and outcome monitoring. However, the accuracy of 2D-US is limited, especially in cases of organ enlargements and deformations. Software-based “stitching” of separately acquired 3D-US data revealed precise volume determination in thyroid phantoms. The purpose of this study is to investigate the feasibility and accuracy of 3D-US stitching in patients with structural thyroid disease. A total of 31 patients from the clinical routine were involved, receiving conventional 2D-US (conUS), sensor-navigated 3D-US (3DsnUS), mechanically-swept 3D-US (3DmsUS), and I-124-PET/CT as reference standard. Regarding 3DsnUS and 3DmsUS, separately acquired 3D-US images (per thyroid lobe) were merged to one comprehensive data set. Subsequently, anatomical correctness of the stitching process was analysed via secondary image fusion with the I-124-PET images. Volumetric determinations were conducted by the ellipsoid model (EM) on conUS and CT, and manually drawn segmental contouring (MC) on 3DsnUS, 3DmsUS, CT, and I-124-PET/CT. Mean volume of the thyroid glands was 44.1 ± 25.8 mL (I-124-PET-MC = reference). Highly significant correlations (all p < 0.0001) were observed for conUS-EM (r = 0.892), 3DsnUS-MC (r = 0.988), 3DmsUS-MC (r = 0.978), CT-EM (0.956), and CT-MC (0.986), respectively. The mean volume differences (standard deviations, limits of agreement) in comparison with the reference were −10.50 mL (±11.56 mL, −33.62 to 12.24), −3.74 mL (±3.74 mL, −11.39 to 3.78), and 0.62 mL (±4.79 mL, −8.78 to 10.01) for conUS-EM, 3DsnUS-MC, and 3DmsUS-MC, respectively. Stitched 3D-US data sets of the thyroid enable accurate volumetric determination even in enlarged and deformed organs. The main limitation of high time expenditure may be overcome by artificial intelligence approaches.