Segmentation of common and internal carotid arteries from 3D ultrasound images based on adaptive triple loss

Objective: Vessel-wall-volume (VWV) and localized vessel-wall-thickness (VWT) measured from 3D ultrasound (US) carotid images are sensitive to anti-atherosclerotic effects of medical/dietary treatments. VWV VWT measurements require the lumen-intima (LIB) media-adventitia boundaries (MAB) at common internal arteries (CCA ICA). However, most existing segmentation techniques were capable automating only CCA segmentation. An approach segmenting MAB LIB ICA was required accelerate quantification. Methods: Segmentation for performed independently using proposed two-channel U-Net, which driven by a novel loss function known as adaptive triple Dice (ADTL). A test-time augmentation (TTA) is used, in three times based on axial its flipped versions; final generated pixel-wise majority voting. Results: Experiments involving 224 3DUS volumes produce Dice-similarity-coefficient (DSC) 95.1%$\pm$4.1% 91.6%$\pm$6.6% LIB, CCA, respectively, 94.2%$\pm$3.3% 89.0%$\pm$8.1% ICA, respectively. TTA ATDL contributed statistically significant improvement all except ICA. The total time segment entire volume (CCA+ICA) 1.4s. Conclusion: U-Net with ADTL can accurately efficiently volume. Significance: Our has potential transition atherosclerosis clinical research.