Weakly-supervised segmentation using sparse single point annotations for lumen and wall of carotid arteries in 3D MRI

Background and objective: Segmentation of the carotid artery is a crucial step in planning therapy for atherosclerosis. Manual annotation is a time-consuming and labor-intensive process, and there is a need to reduce this effort. Methods: We propose a weakly supervised segmentation method using only a few annotated axial slices, each with a single-point annotation from 3D magnetic resonance imaging for the lumen and wall of the carotid artery. The proposed method contains three loss functions designed to (1) locate the center point of the vessel, (2) constrain the range of the vessel radius using prior information implemented with spatial maps, and (3) encourage similar segmentation results in adjacent slices. Both the lumen (inner structure) and wall (outer structure) can be segmented by adjusting the range of plausible radii. Results: Experimental evaluations on the COSMOS2022 dataset show that our method achieved similar performance results (DSC_lu 0.821 lumen, DSC_wa 0.841 wall) to those of fully supervised methods with dense annotations (DSC_lu 0.814-0.857, DSC_wa 0.832-0.875). Similar trends were observed on an independent Harvard dataset. Conclusion: Our proposed method demonstrated effective segmentation of crucial arteries, internal carotid artery, external carotid artery, and common carotid artery in atherosclerosis. We anticipate that this efficient approach utilizing single-point annotation will contribute to the effective management of carotid atherosclerosis. Our code is available at https://github.com/jongdory/CASCA.