Evaluating Stenosis Detection with Grounding DINO, YOLO, and DINO-DETR

TL;DR

This work addresses automated stenosis detection in X-ray coronary angiography by benchmarking CNN-based YOLO and transformer-based DINO-DETR and Grounding DINO on the ARCADE dataset within the MMDetection framework, using COCO-style metrics ($IoU$, $AP$, $AR$, $mAP$) to compare performance. Transformer-based detectors generally deliver higher precision and recall across a range of IoU thresholds, with DINO-DETR excelling in consistent detection and Grounding DINO performing well at moderate overlaps, while YOLO emphasizes speed with some accuracy trade-offs. The study also notes ARCADE-specific annotation inconsistencies and software-compatibility challenges, highlighting data quality and tooling as key factors for reproducible evaluation. Overall, the results inform model selection for automated CAD diagnostics, suggest post-processing and data-cleaning as future directions, and point to opportunities in deeper transformer backbones, deformable attention, and domain-specific augmentations to enable robust clinical deployment.

Abstract

Detecting stenosis in coronary angiography is vital for diagnosing and managing cardiovascular diseases. This study evaluates the performance of state-of-the-art object detection models on the ARCADE dataset using the MMDetection framework. The models are assessed using COCO evaluation metrics, including Intersection over Union (IoU), Average Precision (AP), and Average Recall (AR). Results indicate variations in detection accuracy across different models, attributed to differences in algorithmic design, transformer-based vs. convolutional architectures. Additionally, several challenges were encountered during implementation, such as compatibility issues between PyTorch, CUDA, and MMDetection, as well as dataset inconsistencies in ARCADE. The findings provide insights into model selection for stenosis detection and highlight areas for further improvement in deep learning-based coronary artery disease diagnosis.

Figures (5)

• Figure 1: Sample images from the ARCADE dataset with annotations for stenosis detection. The red bounding boxes indicate regions of arterial narrowing.
• Figure 2: Qualitative comparison of original test images (first column) with detection results from DINO-DETR, Grounding DINO, and YOLO.
• Figure 3: Architectural diagram of DINO-DETR, adapted from b6.
• Figure 4: Architectural diagram of Grounding DINO, adapted from b7.
• Figure 5: Architectural diagram of YOLO, adapted from b5.