Recent years have brought increasing development of artificial intelligence (AI)-based approaches aimed at optimizing clinical decision-making in vascular medicine. Models increasingly facilitate diagnostic interpretation, treatment planning, and prognosis prediction, particularly in complex cases of peripheral artery disease (PAD). In parallel, a major effort to integrate deep learning methods and imaging modalities has led to new opportunities for using anatomical data to enhance diagnostic capabilities. Among these modalities, magnetic resonance imaging (MRI) is gaining popularity as a noninvasive, nonionizing imaging tool with excellent soft tissue contrast and plaque characterization capabilities. Unlike conventional imaging, MRI can provide information on vascular composition—eg, soft tissue, fibrosis, and calcification—without exposing PAD patients to ionizing radiation and thus is an especially desirable modality for serial monitoring and long-term follow-up of PAD. In this review, we present an overview of the current state of AI integration with MRI in vascular imaging, with a particular focus on PAD. We also introduce several contributions from our research group, with our uniquely designed MRI protocol that incorporates ultrashort echo time and multicontrast sequences to improve vascular tissue characterization in PAD. These high-resolution datasets have served as the foundation for the development of deep learning-based classification systems, including both supervised and unsupervised models. Specifically, we describe how variational autoencoders and hybrid scoring systems have been used to phenotype PAD lesions and predict procedural factors such as guidewire crossability. Our work represents a step toward more objective, reproducible, and clinically interpretable tools that bridge high-fidelity imaging with scalable, AI-driven analysis.

Background: Lower limb arterial calcification (LLAC) is a robust imaging biomarker of peripheral artery disease (PAD) severity. Vitamin K antagonists are presumed to accelerate cardiovascular calcification. Direct oral anticoagulants (DOACs) may influence vascular calcification differently, but lower limb data are limited. Methods: We performed a single-center retrospective cross-sectional study comparing LLAC on clinically acquired non-contrast CT between DOAC users and controls without anticoagulation. Patients were propensity score-matched 1:1 (48 DOAC vs. 48 control; n = 96) using baseline clinical covariates. Associations between LLAC scores and perioperative or cardiovascular events were assessed. Segment-specific LLAC was quantified on non-contrast CT and normalized for arterial segment length. A prespecified exposure–duration sensitivity analysis compared the outcomes in patients with ≥5 years of continuous DOAC therapy (n = 22) versus matched controls. Results: In the matched cohort, total LLAC scores did not differ significantly between DOAC and control groups (infrarenal aorta: median 7596.0 vs. 8637.0 (p = 0.487), iliac segment: median 5689.5 vs. 5193.5 (p = 0.602). However, in patients with ≥5 years of DOAC use, LLAC scores were significantly lower in proximal segments: infrarenal aorta median 5593.5 vs. 11,185.0 (p = 0.001997) and iliac arteries 5624.5 vs. 11,501.0 (p = 0.001867)). Higher LLAC was associated with major adverse cardiovascular events (such as myocardial infarction, stroke, or significant bleeding) in controls (p = 0.0023) but not in DOAC-treated patients. Conclusions: In this propensity-matched, cross-sectional CT study, long-term DOAC exposure was associated with lower proximal LLAC scores in a small duration-defined subgroup, while the primary matched analysis showed no overall difference in total LLAC scores. Because baseline (pre-DOAC) imaging was unavailable and residual confounding/survivor bias is possible, these findings should be considered hypothesis-generating and require prospective validation. The cohort reflected a mixed lower-extremity vascular population rather than exclusively classic chronic atherosclerotic PAD, which may limit biological interpretation and generalizability.