Abstract
Background/Objectives: To develop and validate deep learning-based models for detecting significant steno-occlusion (SSO)-defined as luminal narrowing greater than 50%-of the femoropopliteal arteries using maximum intensity projection (MIP) images from lower extremity CT angiography (CTA). Methods: This retrospective study utilized MIP images of lower extremity CTA performed between January 2021 and December 2023 for internal model development. Deep learning-based models were developed sequentially to diagnose SSO: screening with single anteroposterior image, followed by four-segment rotational analysis that divided each femoropopliteal artery into four segments and incorporated multi-angle images. Given the cropped images and the shape of stenosis, models were trained to classify the presence of SSO. A temporal validation dataset comprised MIP images from lower extremity CTA performed between January and June 2024. Results: In total, 56,496 segment images from 642 patients (mean age: 68.2 ± 13.5 years; 472 men) were included in the internal dataset. In the single-image analysis, RDNet achieved the highest mean AUC of 0.886 for SSO detection. In the four-segment rotational analysis, RDNet also demonstrated the highest mean AUC, reaching 0.964 in both half-set and full-set approaches. While RDNet recorded the highest mean AUC, all other models showed improved AUCs as the number of input images increased (p < 0.05). In the temporal validation dataset, RDNet again achieved the highest mean AUC (0.959) in the half-set analysis. Conclusions: The deep learning-based model, particularly RDNet, demonstrated excellent performance in detecting SSO of peripheral arteries on MIP images from lower extremity CTA.