Assessing the transferability of micro-computed tomography (CT)-based plaque segmentation to clinical CT using...

Noninvasive plaque characterization remains limited in peripheral arterial obstructive disease (PAOD), whereas ex vivo micro-computed tomography (micro-CT) enables quasihistological assessment. This study evaluates whether plaque-related structural information-defined as compositional and morphological plaque features learned from micro-CT segmentation-can be transferred to clinical computed tomography (CT) using a super-resolution (SR) framework. Popliteal artery segments from six patients with peripheral arterial obstructive disease were analyzed using micro-CT and histology. Annotated micro-CT images were used to train convolutional neural networks for plaque segmentation. Low-resolution clinical CT images were upsampled using a Laplacian pyramid SR approach, and segmentation models were applied without retraining. Performance was assessed using Dice scores on held-out micro-CT test data, and segmentation outputs on SR-CT images were qualitatively evaluated. Segmentation of calcified plaque components on micro-CT test images yielded Dice scores ranging from 0.58 to 0.67, indicating low-to-intermediate agreement. When applied to SR-CT images, segmentation revealed nonrandom identification of calcified structures in selected image sequences, with marked heterogeneity across slices. SR enables exploratory assessment of plaque information transfer but does not overcome the fundamental resolution gap. These findings define the current limits of CT-based plaque characterization and provide a framework for evaluating future imaging technologies.

Accurate plaque characterization remains a major unmet need in peripheral arterial disease, where treatment planning is largely guided by lesion length and stenosis severity rather than plaque composition. This study proposes using micro-computed tomography (micro-CT) as a histology-informed reference to evaluate how much plaque-related information can be transferred to clinical CT. By explicitly defining current limitations, our findings caution against premature clinical application while informing future developments in advanced CT technologies, multimodal imaging, and artificial intelligence.