Objectives: MRI is a promising tool for identifying tissue properties and enhancing the precision of endovascular peripheral artery disease (PAD) treatment. However, the interpretation of images might pose a challenge for untrained experts. Hypothesis: Unsupervised artificial intelligence (AI) algorithms can assist in interpreting MRI images and classifying lesion components. Aims: To confirm the feasibility of a custom-made variational autoencoder (VAE) in evaluating lesion crossability through histological validation. Methods: 9.4T MRI images (150-micron isotropic resolution) were acquired on 20 chronic total occlusion PAD lesions using T2-weighted (T2w) and Ultrashort Echo Time (UTE) contrasts. Pseudo-colors (T2w: red, UTE: green) were created to enhance tissue visualization (Fig1). A VAE algorithm classified axial images into categories based on the lumen being occluded with hard or soft tissue. Lesions underwent histologic analysis, slides were matched with the corresponding VAE output images. Two readers assessed lesion crossability (non-crossable if >50% of lumen was occluded with hard tissue) on both VAE-generated images and histologic slides. Results: MRI produced 4013 psuedo-color images, and VAE successfully separated axial sections based on plaque composition. Histologic analysis yielded 91 cross-sections. The consensus reading identified 34% (n=31) of VAE images and 25% (n=23) of histologic slides as non-crossable, with a significant association between the histology-based and AI-based crossability (p<.001). Inter-rater reliability was good for both VAE-based and histology-based crossability assessments (intralass correlation coefficient (ICC)= 0.76 and 0.80, respectively). Conclusion: Our AI algorithm demonstrated a strong association with histologic validation, suggesting potential for enhancing image interpretation, and possibly contributing to improved planning for endovascular interventions.