Deep learning-based classification of superficial femoral arterial lesions: a pilot study

Abstract

Purpose

This study evaluated the feasibility of detecting peripheral arterial lesions in plane-reconstructed lower extremity CT angiograms using object detection algorithms.

Methods

We retrospectively collected 1,241 contrast-enhanced lower extremity CT images from patients with peripheral arterial disease. One-stage (YOLOv5: v5s, v5m, v5l, v5x) and 2-stage (Faster R-CNN) detectors were used to classify stent, stenosis, and occlusion. A one-by-one comparison between manual test image annotations and algorithmic detection results was conducted to evaluate model performance and errors. Performance was evaluated by mean average precision (mAP@.5) and precision-recall curves.

Results

Among YOLOv5 models, v5l showed the highest overall accuracy (77% mAP@.5). While stent classification was excellent (≥ 96.9% mAP@.5 in YOLOv5 and 99.8% in Faster R-CNN), classification accuracies for stenosis (53.8%–58.7% in YOLOv5 vs. 37.2% in Faster R-CNN) and occlusion (69%–80.9% in YOLOv5 vs. 67.7% in Faster R-CNN) were moderate. Stenosis was frequently missed, resulting in high false-negative rates. Occlusions at arterial bifurcations were often not detected, and stent edges were misclassified as occlusions. Overfitting emerged in some YOLOv5 models beyond 75 epochs.

Conclusion

This pilot study supports the feasibility of applying object detection algorithms as a preliminary step toward developing clinical decision support tools for peripheral arterial disease. Further refinements, including additional training data and more granular lesion annotation, are essential for improved classification of stenosis and occlusion.

INTRODUCTION

Peripheral arterial disease (PAD) of the lower extremities stems from atherosclerosis, causing ischemic pain, functional impairment, and ulcers, gangrene, or limb loss in advanced cases [1,2,3]. The prevalence of PAD continues to rise globally, especially in regions with higher sociodemographic indices [4], predominantly affecting the superficial femoral and popliteal arteries and manifesting as calf claudication [5,6,7]. Initial management focuses on risk-factor control through lifestyle modifications, rigorous management of hypertension, diabetes, and dyslipidemia, along with antiplatelet therapy [8]. For refractory or advanced cases, intervention may be necessary through endovascular procedures, open surgical treatment, or hybrid approaches [9,10]. Diagnostic evaluation starts with a focused history, physical examination, and ankle-brachial index (ABI). When imaging is needed, CT angiography (CTA) is widely used in Korea for its high accuracy, reproducibility, 3D reconstruction capabilities, brief examination time, and cost-effectiveness [11,12].

Although PAD prevalence rises markedly with age, it remains less recognized than myocardial infarction or stroke [2,5]. Prolonged asymptomatic phases can lead to diagnoses only at advanced stages, especially when patients present with nonspecific symptoms and may delay testing [2,13,14,15]. In particular, waiting for an expert vascular radiologist’s interpretation often prolongs treatment decisions, underscoring the need for faster image analysis and more efficient clinical workflows. Consequently, recent research has employed natural language processing and machine learning to identify PAD patients early, predict prognosis, and reduce the workload on vascular specialists through clinical decision support systems (CDSS) [16].

Lower extremity CTA performed for PAD patients often comprises extensive imaging across multiple series, with numerous vascular lesions that may be detected, each of which can change over time. This creates a significant burden in generating an accurate report. Therefore, rather than merely classifying images, object detection algorithms capable of detecting each vascular lesion individually, along with performing localization and segmentation, may be particularly useful in PAD. The challenge, however, lies in the fact that precisely annotating and training on every CT image requires substantial resources and cost.

The utility of algorithm-based approaches has been consistently demonstrated not only in simulated settings but also in real-world practice, where they have shown potential to improve workload efficiency [17,18]. Notably, a recent study applying an object detection algorithm to vascular imaging for abdominal aortic aneurysm screening reported that even non-expert examiners achieved accuracy comparable to that of experienced operators [19]. Therefore, we aimed to introduce an object detection algorithm to identify and classify vascular lesions from CT scans. However, since CT scans contain extensive information that is scattered away multiple series, identifying each lesion is a time-consuming job. Therefore, in this study, we aimed to leverage maximum intensity projection (MIP) images of superficial femoral artery (SFA), which is a reconstructed 2-dimensional (2D) representation from 3-dimensional (3D) lower extremity vasculature, to reduce the annotation and computational burden while maximizing the advantages of object detection algorithms in identifying individual lesions.

We selected Faster R-CNN and YOLOv5 as representative 2-stage and 1-stage object detection models, respectively. Two-stage models such as Faster R-CNN generally achieve higher accuracy but require more computational resources and longer inference times. In contrast, YOLOv5 offers markedly faster inference and greater computational efficiency, albeit with a slight trade-off in accuracy. Faster R-CNN was thus chosen for its robust localization capabilities suited to precise lesion annotation, while YOLOv5 was selected for its ability to detect small and variable-sized lesions in real time. Comparing these complementary approaches allowed us to assess their feasibility for clinical application in peripheral arterial lesion detection.

Owing to their speed, accuracy, and ability to generalize across different conditions, object detection algorithms show great promise as tools for medical image-based CDSS. Building on this foundation, we applied these algorithms to MIP images reconstructed from 3D CT scans of the SFA. Through a comparison of a 1-stage model (YOLOv5) and a 2-stage model (Faster R-CNN), this study explores the feasibility of automatically detecting stenosis, occlusions, and previously inserted stents. This study serves as an exploratory step toward assessing whether object detection algorithms can eventually contribute to decision support systems in vascular imaging.

METHODS

This is a retrospective study evaluating the feasibility of deep learning-based object detection algorithms for detecting stenosis, occlusion, and stent lesions in the SFA using MIP images derived from lower extremity CTA.

Study population and data collection

The study population comprised contrast-enhanced lower extremity artery 3D CT image data obtained from patients who either presented with symptoms suggestive of lower extremity PAD (LE-PAD) or had already been diagnosed with LE-PAD, and who underwent imaging at our hospital between January 1, 2018, and April 30, 2021. These images were obtained from 3 patient groups: (1) individuals presenting with symptoms suggestive of vascular claudication, (2) those referred to our vascular surgery department after an ABI of 0.9 or lower was identified during routine health screening or evaluation for other conditions, and (3) patients diagnosed with critical limb ischemia who underwent follow-up CT imaging for treatment monitoring. All CT scans were performed using a SOMATOM Force system (Siemens) with a 1 mm slice thickness. CT images were collected from all patients regardless of sex, age, or underlying comorbidities. A single vascular surgeon then reviewed the images and excluded those deemed unsuitable for model training, such as images with poor resolution or excessive artifacts.

After reconstructing acquired CT images into 3 dimensions, MIP images were generated from a frontal viewpoint. MIP is a visualization technique that converts 3D imaging data (e.g., CT, MRI, single-photon emission computed tomography) into a 2D image by identifying the voxel with the highest intensity along a defined viewing direction and projecting it onto the 2D plane. This process selects only the maximum-intensity voxel for each corresponding position in the 2D image, thereby making high-intensity structures—such as contrast-enhanced vessels, calcifications, and bones—stand out compared to surrounding tissues with lower intensity. The MIP image reconstruction was automatically performed and stored in the Picture Archiving and Communication System (PACS). From PACS, anteroposterior images showing the entire SFA were extracted in a de-identified form without any patient-identifying information. The images were converted to JPEG format for storage on a local computer, and no additional preprocessing or data augmentation was performed.

Classification of vascular lesions with an image classification algorithm

Image annotation was conducted by a vascular specialist to classify each type of lesion and localize its position. The annotation program used was LabelMe (Python 3.6 version). A rectangular bounding box was drawn around each lesion, and an appropriate class name was assigned. Lesions were classified and annotated into 3 categories: stent, stenosis, and occlusion.

Stenotic lesions were annotated where the vessel lumen showed focal or diffuse narrowing compared to adjacent normal segments while maintaining continuity with proximal and distal segments. In accordance with existing literature, a lesion was considered stenotic when the vessel diameter was reduced by 50% or more relative to the adjacent normal reference segment [20]. The bounding box was drawn from the point where vessel wall irregularity or luminal narrowing first appeared to the point where the vessel returned to its original caliber. When multiple stenotic segments occurred in sequence without a clearly normal intersegment, they were annotated as a single continuous lesion.

Occlusive lesions were annotated where the vessel showed an abrupt interruption of continuity, appearing on MIP images as a sudden narrowing or disappearance of the contrast-enhanced arterial lumen. To help the model learn this interruption pattern more effectively, the bounding boxes were drawn to encompass not only the visibly occluded segment but also a portion of the proximal patent vessel and adjacent background tissue. In general, a margin of a few millimeters proximal and distal to the occlusion site was included to better capture the anatomical transition into the occlusion and support accurate recognition by the model.

Stents were annotated based on their radiological features including high radiopacity, tubular structure, and mesh-like pattern. They appeared as bright linear or cylindrical structures contrasting sharply with the surrounding tissues. Bounding boxes were drawn to encompass the entire visible stent, including its full length and lateral extent. Examples of lesion annotations from the representative test set are provided in Supplementary Fig. 1 for reference.

Dataset allocation and model training

As a pilot study with limited data, we prioritized minimizing overfitting by allocating most images to training and validation sets, while the test set was limited to 6 representative images selected by a vascular specialist who performed the annotation. This allowed for targeted, case-by-case evaluation of the model’s detection and classification performance across diverse lesion types and anatomical locations. Those images include diverse presentations of all 3 lesion classes (stenosis, occlusion, and stent) located in various segments of the SFA and of varying sizes. The test set was entirely excluded from training and was used exclusively for final performance evaluation. All remaining images were assigned as training images, and approximately 10% of these were further designated as the validation set for each training cycle through computerized randomization.

Additionally, a subset of normal images was included in the training data as background to prevent overfitting to lesion-containing images and to minimize false positives (i.e., classifying normal vessels as diseased). To mitigate overfitting, we included 50 background-only images (about 5% of the training set) in line with object detection practices recommending 0%–10%. Given the class imbalance among lesion types, this proportion was chosen to balance the risks of false positives and false negatives by preventing overfitting to either lesion-containing or background images [21,22].

This study employed 2 CNN-based object detection algorithms: YOLOv5 and Faster R-CNN. For YOLOv5, we evaluated all 4 network sizes (v5s, v5m, v5l, and v5x) to compare their performance. Each model detected lesion objects within images and calculated the probability that each object belonged to 1 of the 3 lesion classes. After classification, the predicted class was compared to the ground-truth label. A prediction was considered correct if the predicted bounding box achieved an intersection over union (IoU) of at least 0.5 with the ground-truth b ounding box.

$$\text{Intersection over union (IoU)}=\frac{\text{Area of overlap}}{\text{Area of union}}$$

Using this criterion, model validation was performed during each training epoch by measuring objectness loss and classification loss for both the training and validation sets. After 100 training epochs, the final evaluation was conducted using the test set. The performance metrics included a confusion matrix, precision-recall curves, and mean average precision at an IoU threshold of 0.5 (mAP@0.5).

Because YOLOv5 and Faster R-CNN demand substantial computational power, we utilized the NVIDIA Quadro RTX 5000 GPU (16 GB). For YOLO v5, optimizer hyperparameters were set to a learning rate of 0.01, momentum of 0.937, and weight decay of 0.0005. Model hyperparameters were set to an image size of 512, a confidence threshold of 0.001, and a nonmax suppression IoU threshold of 0.6. All hyperparameters were kept at the algorithm’s default settings without further tuning.

Code availability

The deep learning source code used in this study is publicly available in a GitHub repository. The code includes all major components used for the detection and classification of SFA lesions from MIP images utilizing YOLOv5 and Faster R-CNN.

The full code for the YOLOv5 implementation is publicly available at: https://github.com/YJWoo6514/DL-for-SFA_YOLO. The Faster R-CNN implementation is available at: https://github.com/YJWoo6514/DL-for-SFA_FasterRCNN.

Study ethics

The study protocol was reviewed and approved by the Institutional Review Board (IRB) of The Catholic University of Korea, Seoul St. Mary’s Hospital (No. IRB#KC21RASI0745). Given that all patient data were deidentified prior to analysis, the IRB determined that obtaining informed consent from participants was not required, as the study posed minimal risk to the rights and welfare of the subjects.

RESULTS

Demographics of image of interest

During the study period, 2,318 CT lower extremity artery scans were performed at our institution. Of these, 1,077 scans were excluded due to poor resolution (n = 698), motion or metal artifacts from prior surgical interventions (n = 351), or incomplete series preventing proper MIP image generation (n = 28). This left 1,241 scans for inclusion in the study. From these, a board-certified vascular specialist selected 6 representative lesion-containing images featuring all 3 lesion types (stenosis, occlusion, and stent) for the test set. The remaining 1,235 lesion-containing images were divided using computerized randomization into training (n = 1,110) and validation sets (n = 125) at a 9:1 ratio. To reduce false positives and enhance model generalizability, we added 50 background-only images to the training set, bringing it to 1,160 images. In total, we annotated 2,736 lesions across all datasets. The distribution of images and lesion types across training, validation, and test sets is summarized in Fig. 1 and Table 1.

Fig. 1. CONSORT (Consolidated Standards of Reporting Trials)-style flow diagram of image selection and dataset allocation. Six images were pre-selected for the test sets to assess case-wise detection, and the remaining images were randomly assigned to training and validation sets via computerized randomization. Background-only images were only included in the training set to prevent overfitting and reduce false positives. CTA, CT angiography; LE-PAD, lower extremity-peripheral artery disease.

Table 1. Number of lesions designated to the training, validation, and test sets.

Detection and classification of vascular lesions by YOLOv5

Fig. 2 shows an example of image that has undergone the detection and classification of vascular lesions by YOLOv5l. Among YOLOv5 models tested, the v5l model exhibited the highest overall accuracy at 77.7% mAP@.5, making it the top performer overall. With respect to individual lesion classes, the stent classification accuracy was the highest in both v5s and v5x models at 99% mAP@.5, while occlusion classification accuracy was the highest in the v5l model at 80.9% mAP@.5. For stenosis, the v5x model achieved the highest accuracy at 58.7%. These findings indicate that classification performances varied slightly by lesion class and model type. However, while all models performed exceptionally well in distinguishing stents, their ability to differentiate stenoses and occlusions remained moderate, ultimately contributing to a reduction in overall accuracy.

Fig. 2. Examples of 3-dimensional–reconstructed lower extremity artery CT images after identification and classification of vascular lesions by YOLOv5l.

Changes in recall and precision according to the confidence threshold were depicted using a precision-recall curve. Overall accuracy across the models ranged from 75% to 77.7% mAP@.5, showing no clear-cut superiority of one model over another. In general, all models demonstrated very high accuracies for stents (96.9%–99.2% mAP@.5). However, the accuracies for occlusion ranged from 69% to 80.9%, and the accuracies for stenosis remained comparatively low at 53.8%–58.7%, negatively affecting the overall accuracy. Table 2 presents a summary of precision, recall, and mAP@.5 values for each model by class, and Fig. 3 visualizes these results using the precision-recall curve.

Table 2. Precision, recall, and mAP@.5 of each YOLOv5 model for the classification of stenosis, occlusion, and stent.

Fig. 3. Precision-recall curves of each YOLOv5 model for the classification of stenosis, occlusion, and stent. (A) YOLO v5s, (B) YOLO v5m, (C) YOLO v5l, and (D) YOLO v5x.

We also examined how classification and objectness performances varied with the number of epochs for each YOLO v5 model. In general, all models displayed a decreasing trend in classification loss and objectness loss on both training and validation data as the number of epochs increased. However, for v5m, v5l, and v5x models (with the exception of v5s), objectness loss showed a slight increase once the epoch count exceeded 75. Fig. 4 provides a visualization of training and validation losses for each model across epochs.

Fig. 4. Loss graphs of each model according to epoch. (A) YOLO v5s, (B) YOLO v5m, (C) YOLO v5l, and (D) YOLO v5x.

A one-by-one comparison between manual annotations (Supplementary Fig. 1) and YOLOv5l detection results (Fig. 2) for 6 test images was conducted. Out of 15 annotated stenoses, 9 were correctly detected and 6 were missed. For occlusions, 2 out of 4 were detected, with 4 false positives. All 5 stents were correctly detected, with only 1 false positive. The false positives for occlusion primarily occurred where vessels lost patency within pre-inserted stents, showing no continuation beyond the stent’s distal end. Although we did not annotate these cases as occlusions, the algorithm misinterpreted the vessel discontinuity at the stent’s distal edge as occlusions. Table 3 provides a detailed summary of false positive and false negative results for each test image.

Table 3. Per-image summary of false positive and negative detections by YOLOv5l for stenosis, occlusion, and stent in the test dataset.

Detection and classification of vascular lesions by Faster R-CNN

Fig. 5 shows an example of image that has undergone detection and classification of vascular lesions by Faster R-CNN. Using Faster R-CNN on the same set of images, stenosis accuracy was 37.2% mAP@.5, occlusion accuracy was 67.7% mAP@.5, and stent accuracy was 99.8% mAP@.5, resulting in an overall accuracy of 68.2% mAP@.5. Although Faster R-CNN achieved slightly higher accuracy for the stent class than the 1-stage object detector methods, its accuracy for stenosis was substantially lower. This reduction in stenosis accuracy led to a lower overall accuracy than YOLOv5. Table 4 shows the comparison of precision and recall between YOLOv5l and Faster R-CNN.

Fig. 5. Examples of 3-dimensional–reconstructed lower extremity artery CT images after identification and classification of vascular lesions by Faster R-CNN.

Table 4. Comparison of precision, recall, mAP@.5 of YOLO v5l and Faster R-CNN for the classification of stenosis, occlusion, and stent.

Fig. 6 illustrates classifier loss, objectness loss, and box regression loss of Faster R-CNN across epochs. Similar to YOLOv5, as the number of epochs increased, both objectness and classification loss showed a decreasing trend. Precision and recall for stenosis, occlusion, and stent were visualized using precision-recall curves.

Fig. 6. Loss graphs and precision-recall curves of Faster R-CNN for classification of stenosis, occlusion, and stent. (A) Loss graph of Faster R-CNN according to epoch. (B) Precision-recall curve for each classification. Reg, regression; AP, average precision.

A one-by-one comparison between manual annotations (Supplementary Fig. 1) and Faster R-CNN detection results (Fig. 5) for 6 test images was conducted. Out of 15 annotated stenoses, 7 were correctly detected and 8 were missed. For occlusions, 2 out of 4 were detected, with 4 false positives. All 5 stents were correctly detected. The pattern of false positive detections for occlusions was similar to that observed with YOLOv5l. Table 5 provides a detailed summary of false positive and false negative results for each test image.

Table 5. Per-image summary of false positive and negative detections by Faster R-CNN for stenosis, occlusion, and stent in the test dataset.

DISCUSSION

For accurate surveillance of PAD, it is essential to document in detail the lesions identified on CT in the radiology report and to compare them with prior findings. However, in lower extremity CT series, a large number of images are generated, and producing such a report requires substantial time. Therefore, this study aimed to leverage the strengths of object detection algorithms to determine whether multiple lesions could be simultaneously detected and classified from MIP images in PAD patients. The results demonstrated that even in test images containing multiple lesions simultaneously, the models could detect and classify each lesion with considerable accuracy, successfully achieving the anticipated outcomes of adopting object detection algorithms.

These results may be utilized to improve clinical workflows in the following ways. First, the system can highlight suspected lesion areas, informing the radiologist where to focus detailed evaluation, thereby reducing the risk of missed lesions. Second, by comparing these results with prior assessments, it is possible to determine whether each object is newly developed or corresponds to a pre-existing lesion, facilitating follow-up evaluation. Third, although not implemented in this study, if stenosis severity could also be assessed, the temporal progression of each object could be quantitatively analyzed. Such capabilities would substantially reduce the radiologist’s workload while allowing precise documentation of each lesion and reducing inter-observer variability.

Notably, the 1-stage algorithm demonstrated higher accuracy than the 2-stage algorithm, particularly for stenotic lesion assessment. This suggests that the 1-stage algorithm can perform PAD surveillance both rapidly and with high accuracy, indicating strong potential for future clinical application.

However, the performance of both models was largely attributable to easy-to-detect stents, while the detection performance for stenosis and occlusion, which is the primary focus of this study, was lower. Specifically, at the object level, stenosis exhibited many false negatives, while occlusion had many false positives. This is because stents are large lesions with a distinct pattern that is easily differentiated from surrounding structures, whereas other lesions are not.

Therefore, to enhance the clinical utility of the model, detection accuracy for stenosis and occlusion must be improved. Several approaches could be attempted to address this. First, various labeling strategies should be explored. In this study, simple rectangular labeling was used, but adopting polygonal labeling could exclude surrounding soft tissue and include only stenotic features such as vessel diameter reduction or a beaded appearance. For occlusions, to enable the model to learn the pattern of “persistent flow followed by abrupt termination,” various labeling ranges before and after the occlusion should be tested, and the most appropriate strategy should be standardized. This could help detect irregularly shaped occlusion lesions, such as arterial branch occlusions seen in the upper middle and lower right of Supplementary Fig. 1.

Second, incorporating changes in blood flow across the entire lower extremity into object detection and classification could be beneficial. In particular, stenosis and occlusion are often accompanied by features such as collateral circulation development and reduced distal runoff, and leveraging such additional information may aid lesion interpretation.

Finally, when disease progression occurs after stent placement and only the stent remains visible, the proximal and distal segments are frequently misclassified as occlusion. In the future, such cases should either be excluded from the occlusion class or classified separately as in-stent occlusion within the stent object, using a staged classification approach to improve accuracy.

This study has several major limitations. First, due to the limited number of board-certified physicians on our research team, we could not validate the ground-truth labels through consensus annotation by multiple vascular specialists. Although we applied predefined criteria for lesion labeling, ensuring the reliability of ground-truth annotations remains a critical methodological concern, particularly in object detection tasks where bounding box placement plays a central role in model learning and evaluation. In traditional classification tasks, a common strategy is to have 2 independent annotators label the dataset and to resolve discrepancies through consensus or adjudication by a third expert. However, this approach becomes less straightforward in object detection. Since bounding boxes involve continuous spatial coordinates, even when the annotators agree on the presence of a lesion, small differences in box boundaries are unavoidable. As a result, in object detection tasks, every case requires a decision about which bounding box should be accepted as the ground truth. Therefore, we propose that future studies adopt a 2-tiered annotation protocol, wherein one board-certified vascular specialist serves as the primary evaluator and another as a validator. The validator would review the annotations for correctness and consistency, and adjustments to box boundaries could be tracked quantitatively using overlap metrics such as IoU. This framework would enable a more structured approach to verifying annotation reliability and improving dataset consistency.

Second, this study used single-center data with a limited dataset size and allocated few images to the test set. This non-randomized allocation was intentional to meet the goals of a pilot study aimed at determining whether diverse objects could be detected in representative test images and identifying potential issues. However, it limited statistical power for objectively analyzing and comparing model performance, so performance comparisons were restricted to mAP@0.5, which limit its generalizability and granularity. Additionally, all models except YOLOv5s exhibited signs of overfitting after 75 epochs, indicating that more data are needed for adequate training. Therefore, future research should include data collection from multiple centers or increase sample size through data augmentation, and implement methods such as early stopping or dropout to prevent overfitting in deeper architectures.

Third, the MIP images used in this study have the advantage of enabling assessment of all vascular lesions at once, but they inherently lose detail when projecting 3D voxel data onto a 2D plane. In particular, distinguishing between a stent and in-stent restenosis or occlusion is challenging. As noted earlier, if the stent class were subdivided into in-stent occlusion and in-stent restenosis, and an appropriate labeling method for distinguishing them were identified, relatively accurate information could be obtained even from limited data. Similarly, in this study, stenosis severity was not labeled. If the model could first identify stenosis and then classify its severity as a subclass, it would be easier to track changes in each lesion over time during follow-up.

Therefore, the next phase of research should use a larger dataset and allocate sufficient images to the test set to fully train each model and compare their performance. At the same time, to enhance model accuracy, optimal annotation and lesion classification systems should be established so that the model is tailored for PAD. Once such foundational research is completed prior to clinical implementation, it would then be possible to assess the computational resources and response times required, as well as investigate integration into actual clinical workflows and the potential development of a CDSS.

By applying object detection algorithms to MIP images of PAD patients, multiple lesion objects corresponding to stenosis, occlusion, and stent can be identified and classified within a single planar image. The 1-stage method YOLOv5 performed object detection slightly more accurately than the 2-stage model Faster R-CNN in a pre-selected test set, while also being relatively faster. However, when evaluated on an object-count basis rather than an area-based metric, no distinct difference between the models was observed. Before clinical implementation, it is necessary to compare various annotation methods, establish class and subclass hierarchies to describe each lesion more precisely, and adopt 2-tiered annotation to improve classification reliability, followed by exploration of models optimized for PAD.

SUPPLEMENTARY MATERIALS

Supplementary Fig. 1 can be found via https://doi.org/10.4174/astr.2026.110.2.92.

Supplementary Fig. 1

Example of annotated vascular lesions for the test dataset. Boxes colored red, green, and yellow indicate stenosis, occlusion, and stent, respectively.