Deep learning diagnosis of femoropopliteal artery stenosis using MIP images from CT angiography

Sunday, March 30, 2025

3:09 PM – 3:18 PM CT

Location: Music City Center, 205 BC

Presenting Author(s)

Lyo Min Kwon, MD (she/her/hers)

Assistant professor
Hallym University Sacred Heart Hospital, Republic of Korea

Author/Co-author(s)

WH

Wonju Hong, MD

Clinical assistant professor
Hallym University Sacred Heart Hospital, Republic of Korea

Purpose: Peripheral arterial occlusive disease (PAOD) can be diagnosed using lower extremity computed tomography angiography (CTA), assessing stenosis in arteries of lower extremities. We aimed to develop and evaluate a deep learning-based detection algorithm for the preliminary screening and interpretation of significant stenosis in the lower extremity arteries, with a focus on the femoropopliteal artery, based on maximum intensity projection (MIP) images of lower extremity CTA in patients with PAOD.

Materials and Methods: MIP images of lower extremity CTA performed between January 2021 and December 2023 in our institute were included in this retrospective study. To facilitate evaluation of stenosis, femoropopliteal artery was divided into four sections of regions of interest (ROI): proximal, mid, distal femoral artery, and popliteal. Three additional images from different angles in addition to the anteroposterior view were analyzed to address the issue of stenosis not being detected at certain angles. In the preprocessing step, we applied normalization of brightness to ensure consistent brightness across images from different patients, and center-symmetric zero-padding was applied to prevent shade distortion during model input. Given the size of the cropped images and the shape of the stenosis, we used a convolutional neural network model for the classification task of determining the presence of significant stenosis.

Results: Total of 11,904 images were collected from 390 patients. 10,292 images were labeled as negative, defined as less than 50% stenosis, and 1,612 images were labeled as positive, defined as over 50% stenosis. Data were assigned to one of three datasets: 9,556 training, 1,176 validation, and 1,172 test dataset. The trained model achieved an average area under the receiver operating characteristic curve (AUROC) of 0.958 in performance with a standard deviation of 0.006 and an average classification accuracy of 93.1% with a standard deviation of 1.1%.

Conclusion: The model showed excellent performance in detecting significant stenosis of peripheral arteries on MIP images of lower extremity CTA. It is expected to assist radiologists by pre-selecting the presence and location of significant stenosis, reduce the workload of radiologists, shorten the interpretation time, and help ensure timely treatment for patients.