Abstract
BACKGROUND: Simultaneous non-contrast angiography and intraplaque hemorrhage (SNAP) imaging allows multi-contrast MR images with large longitudinal coverage to be acquired in a single scan. With vessel wall boundaries available, vulnerable plaque components can be detected automatically from SNAP images. However, since SNAP imaging has not been previously used for vessel wall identification, vessel wall boundaries were required to be segmented from conventional multi-contrast MRI first before registering to SNAP images. This registration process is not only time-consuming but also prone to errors, potentially compromising subsequent plaque component analysis. PURPOSE: We aim to develop a model that directly segments the vessel wall from SNAP images, thereby eliminating the need for registration from another modality. The proposed model mitigates label noise arising from boundary misregistration. METHODS: The proposed framework has a student-mean teacher architecture, trained in two phases: (i) a warm-up phase, in which the model was trained by well-registered manual segmentations and minimizes Dice loss between predictions and manual labels and (ii) a fine-tuning phase, in which the model was trained by both well-registered and misaligned manual segmentations. This phase involves adversarial training with the fast gradient sign method (FGSM) and a novel surrogate label generator. The generator produced surrogate ground truth boundaries for each misaligned image by computing a weighted sum of the manual segmentation and the pseudo-label, generated through selective hardening of predicted probabilities from the student and mean teacher models. The sum of the adversarial training loss and the Dice loss between the manual and predicted segmentations was minimized to obtain the final segmentation result. During inference, the averaged probability maps from the student and mean teacher models were used to assign voxels to their most probable class. This study utilized 129 image volumes (1474 axial slices), of which 74 volumes (810 axial slices) were well-registered and 55 volumes (664 axial slices) were misaligned. Training involved 110 volumes (55 well-registered and 55 misaligned), while validation and testing sets comprised 9 and 10 well-registered volumes, respectively. RESULTS: The proposed method outperformed existing noisy label learning methods when trained by the same set of misaligned segmentations. Results demonstrate our method's superiority with Dice similarity coefficient of 73.51 ± 11.08% , 90.76 ± 10.21% , and 90.10 ± 10.38% for the vessel wall, lumen, and outer wall segmentations, respectively. CONCLUSION: The proposed segmentation framework effectively integrates noisy and reliable labels to produce accurate vessel wall segmentations directly from SNAP images. By eliminating the need for manual segmentation and inter-modality registration, this approach facilitates more detailed plaque component analysis with reduced interslice distance across a longer arterial segment.