In high-precision medical applications, merely detecting the presence of disease is insufficient; clinicians also require precise information about the location of lesions. The “object-based classification” approach directly addresses this need. In this method, instead of assigning a single label to the entire image, the algorithm aims to draw a bounding box around the stenotic region. The technical process typically involves generating a large number of candidate bounding boxes based on image features, clustering overlapping boxes, and ultimately selecting the most probable box as the final output. Various researchers have employed advanced architectures to tackle this challenge.

In this context, Pang et al. [8] proposed the Stenosis-DetNet network by considering the video nature of angiographic data. This network is an end-to-end model that operates directly on raw video inputs. Their main innovation was the incorporation of prior knowledge regarding the displacement of coronary vessels throughout the cardiac cycle, enabling the model to track and refine the location of stenosis across consecutive frames. They utilized the powerful ResNet architecture as the backbone for deep feature extraction.

Danilov et al. [9] conducted a comprehensive study to identify the most suitable architecture for single-stenosis lesion detection. They evaluated eight different combinations of object detectors (including SSD, Faster R-CNN, and R-FCN) and feature extraction backbones (including MobileNet-v2, ResNet50, Inception-v4, and NASNet). Their findings indicated that, although deeper and more complex models generally achieve higher accuracy, the combination of R-FCN with ResNet101 provides the best balance between detection performance and computational speed for clinical applications.

Li et al. [10] introduced a more complex framework, STQD-Det, which leverages spatial–temporal feature sharing alongside a novel concept referred to as “quantum propagation.” This approach achieved a very high F1-score of 92.39% while processing at 25.08 frames per second. However, the authors also noted that the high computational complexity and reliance on specialized modules remain major barriers to deployment in standard hospital hardware environments.

In many scenarios, precise annotation of individual lesions (i.e., bounding box delineation) is not feasible due to the high level of expertise required from radiologists and the time-consuming nature of the process. Under such conditions, “image-based classification,” in which an entire image or a region of interest is labeled as either “healthy” or “diseased,” offers a more practical alternative. In this context, Moon et al. [11] proposed a two-stage system to address the challenge of low-quality angiographic images. In their approach, the algorithm first automatically identifies and selects frames in which the contrast agent fully fills the vessel. Subsequently, the pre-trained Inception-v3 model—initially trained on the ImageNet dataset—is applied to these selected frames. A key contribution of their work is the integration of Convolutional Block Attention Modules (CBAM) in parallel within the early layers of the network. These attention modules enable the model to focus on vascular structures rather than irrelevant noisy background regions.

One of the major challenges in medical deep learning is the limited availability of annotated data. To address this issue, researchers often divide original images into smaller patches (e.g., 32×32 pixels). Antczak and Liberadzki [12], in addition to releasing a dedicated dataset, proposed an innovative data augmentation strategy to mitigate data scarcity. Assuming that vascular structures exhibit geometric similarity to Bézier curves, they generated synthetic training samples based on this assumption. These artificially generated data were then used to pre-train a lightweight convolutional neural network consisting of only five layers, resulting in improved detection performance.

On the other hand, the use of overly large models on limited medical datasets often leads to overfitting. To address this issue, Ovalle-Magalanz et al. [13] proposed an approach known as “network slicing.” They conducted an extensive search across different layers of VGG16, ResNet50, and Inception-v3 to determine how much of the pretrained knowledge should be retained during transfer learning. Their results indicated that, for coronary artery stenosis detection, Inception-v3 provides optimal performance when only its first three blocks are preserved and fine-tuned, while the remaining layers are retrained to adapt to the new task.

Despite the progress achieved by the aforementioned approaches, each of these methods is associated with specific technical and operational limitations. Image classification-based methods, despite their relative simplicity, suffer from an inherent drawback, namely the lack of precise localization capability. These methods only indicate the presence or absence of disease within an image but fail to provide spatial information about lesion location, which is critical for interventional clinical procedures. In addition, alternative strategies often rely on complex preprocessing pipelines and synthetic data generation techniques (such as Bézier curve modeling) to compensate for data scarcity, which significantly increases computational burden and implementation complexity.

On the other hand, object detection-based methods, while addressing the localization problem, still face the persistent challenge of balancing accuracy and speed. Accurate two-stage models such as Faster R-CNN, due to their heavy architectural design, exhibit long inference times, making their deployment in real-time and intra-procedural angiography systems practically infeasible. In contrast, highly efficient models such as early versions of YOLO often experience reduced accuracy when detecting very small and subtle abnormalities in high-resolution medical images. Moreover, recent advanced and complex frameworks such FPN: in conventional deep neural networks, deeper layers contain strong semantic information but have low spatial resolution, which makes them unsuitable for detecting small objects such as mild stenotic regions. RetinaNet addresses this limitation by employing a FPN built on top of a ResNeXt-101 backbone. The FPN structure consists of a bottom-up pathway and a top-down pathway with lateral connections. This architecture enables the integration of strong semantic features from higher layers with fine-grained spatial details from lower layers. As a result, the model can accurately localize coronary stenosis across multiple scales.

Focal Loss Function: one of the major challenges in object detection training is class imbalance, where most regions in an image correspond to the background (healthy tissue), while only a small portion contains stenotic lesions. Standard loss functions, such as cross-entropy loss, tend to be dominated by a large number of easy negative samples. RetinaNet addresses this issue by introducing the focal loss Function 1. In this formulation, the parameter γ (set to 2 in this study) significantly down-weights easy examples, forcing the model during training to focus more on hard-to-classify samples, particularly stenotic regions.

FL(pt) = −αt(1 − pt)γlog(pt). (1)

This model has been developed with the aim of bridging the gap between high detection accuracy and computational efficiency, making it an ideal choice for clinical environments with limited hardware resources. The architecture of EfficientDet-D3 is illustrated in Fig. 3. The key innovations of this model include:

Bi-directional Feature Pyramid Network (BiFPN): in contrast to the conventional FPN, where information flows in a single direction, EfficientDet employs a more advanced structure known as BiFPN. In this design, nodes that contribute minimally to feature extraction are removed, and additional edges are introduced to facilitate efficient feature transfer across different levels. A key characteristic of BiFPN is its ability to enable bidirectional information flow—both top-down and bottom-up—as well as repeated stacking of these fusion blocks. This results in richer feature representations and more effective integration of multi-level semantic information.

Compound scaling: most object detection models improve accuracy by increasing only one dimension of the network, such as depth or width. In contrast, EfficientDet utilizes a compound scaling strategy in which all dimensions of the network—including input resolution, depth, and width—are scaled simultaneously and in a coordinated manner using a single compound coefficient (ϕ). In this study, the EfficientDet-D3 variant is employed, which is optimized for high-resolution images (896 × 896 pixels). This choice enables the model to capture subtle stenotic patterns in large angiographic images without incurring excessive

In this study, a dataset comprising 8,325 angiographic images from 100 patients with a confirmed diagnosis of single-vessel CAD was used. The data were collected from patients who underwent CAG using Siemens Coroscop and GE Innova systems at the institute for complex cardiovascular disease research in Kemerovo, Russia. All participants provided written informed consent, and the images were acquired in accordance with ethical protocols. The raw images had variable spatial resolutions ranging from 512 × 512 to 1000 × 1000 pixels. Data preparation was performed within an object detection framework, where clinical experts annotated stenotic regions using bounding boxes. A notable characteristic of this dataset is that stenotic lesions typically occupy only a small portion of the vessel, while the majority of the artery remains healthy. This property enables the model to learn not only discriminative features of stenotic regions but also contextual information from healthy vascular tissue within the same frame. For unbiased evaluation, the dataset was split into training (80%), validation (10%), and test (10%) sets. A patient-level, non-random splitting strategy was adopted, ensuring that all images belonging to a single patient were included exclusively in one subset. This approach prevents data leakage between training and testing phases and guarantees that the model is evaluated on entirely unseen patients, thereby providing a realistic assessment of generalization performance. The demographic and clinical characteristics of the dataset are presented in Table 1.

In this study, the mAP is used as the standard quantitative metric to evaluate model performance and the quality of lesion localization. This widely adopted metric in object detection tasks assesses the model’s ability to accurately determine the spatial location of coronary stenosis. To compute mAP, two key components—precision and recall—are first defined as follows:

I. Precision (P): indicates the proportion of positive predictions made by the model that are actually correct, reflecting the accuracy of the predictions.

II. Recall (R): indicates the proportion of true stenotic regions correctly identified by the model, reflecting its sensitivity.

The final evaluation is based on the Intersection over Union (IoU) criterion. In this study, the IoU threshold is set to 0.5. This means that a predicted bounding box is considered a correct detection if its overlap with the ground truth bounding box exceeds 50%. The mAP value is computed as the mean of Average Precision (AP) across all detections, where higher values indicate superior model performance in both discrimination and localization of stenotic regions.

In this study, a carefully designed preprocessing pipeline was implemented to enhance the quality of input data and ensure compatibility with the requirements of deep neural networks. Given the noisy nature of angiographic images and the variability in raw image dimensions, the first step involved image standardization. For the RetinaNet-101 model, all images were resized to a fixed resolution of 640 × 640 pixels to match the network input requirements. In contrast, for the EfficientDet-D3 model, which requires finer vascular detail extraction, images were resized to a higher resolution of 896 × 896 pixels. This increase in resolution enables the model to more accurately capture subtle and clinically significant stenotic patterns that may be lost at lower resolutions.

To improve model generalization and prevent overfitting, targeted data augmentation techniques were applied. Random horizontal flipping was employed to enable the model to recognize stenosis regardless of vessel orientation. In addition, random cropping was used to extract smaller image regions, thereby encouraging the network to focus on local features and small-scale stenotic structures. Specifically for the EfficientDet model, a more advanced augmentation strategy involving random scaling, cropping, and padding-to-square transformation was applied to enhance robustness under varying spatial scales and configurations, reflecting real-world clinical variability.

In the final stage of data preparation, the dataset format was optimized for computational efficiency. Raw images and their corresponding XML annotation files were converted into the binary TFRecord format. TFRecord files store data as serialized binary sequences, significantly reducing Input/Output (I/O) bottlenecks during training. The use of this unified format eliminated the overhead associated with reading large numbers of individual files from directories and maximized data throughput to the Graphics Processing Unit (GPU) during model training.

In this study, the training and evaluation process was implemented using the TensorFlow Object Detection Application Programming Interface (API) (version 2.13) and the Python programming language (version 3.9). To accelerate convergence and leverage previously learned visual representations, transfer learning was employed. Accordingly, the initial weights of the models were pre-trained on the Common Objects in Context (COCO) dataset and subsequently fine-tuned using the angiographic dataset of this study.

Due to differences in architectural complexity and computational requirements, two separate hardware environments were utilized. The RetinaNet-101 model, which operates on standard 640 × 640 pixel inputs, was trained on a local workstation equipped with an NVIDIA GeForce RTX 3080 Ti GPU (12 GB VRAM) and 16 GB of RAM. This configuration was sufficient to handle the computational demands of the FPN and ResNet backbone.

In contrast, the EfficientDet-D3 model, which processes high-resolution inputs (896 × 896 pixels) and employs the more complex BiFPN architecture, required significantly higher memory resources. Therefore, training was conducted on the Google Colab cloud platform using an NVIDIA A100 GPU (40 GB VRAM) and 83 GB of system memory. This hardware separation ensured that each model was trained under optimal conditions without memory bottlenecks, thereby maximizing computational efficiency. A summary of hyperparameter settings and hardware specifications is provided in Table 2.

In this section, the performance of the proposed models (RetinaNet and EfficientDet-D3) is evaluated in terms of accuracy and speed and compared with state-of-the-art methods. The primary metric used for assessing localization accuracy is mAP with an IoU threshold of 0.5.

Table 3 presents the results of evaluating the proposed models in comparison with other recent studies. As shown, the EfficientDet-D3 model achieves the highest accuracy among all evaluated methods, with a mAP of 96.6%. This model outperforms strong existing approaches, including the method proposed by Danilov et al. [9] with an mAP of 95.4% and the complex quantum-based framework of Li et al. [10] with a score of 92.39%. This superior performance can be attributed to the BiFPN architecture and the use of higher-resolution input images (896 × 896), which enable more precise delineation of stenotic boundaries. The RetinaNet-101 model also demonstrates highly competitive performance, achieving an mAP of 93.2%, indicating its robustness in handling the challenges of angiographic imaging.

In addition to accuracy, processing speed is of critical importance for real-time clinical applications. Table 4 reports the inference time of both models on the test set. RetinaNet achieves a faster average inference time of 57 ms per frame compared to EfficientDet-D3, which requires 95 ms. Although EfficientDet-D3 is slower, its significantly higher accuracy (96.6%) makes it a more suitable option for high-stakes diagnostic scenarios where accuracy is the primary priority. Importantly, both models operate within an acceptable time range for intraoperative and clinical use.

A visual inspection of the model outputs demonstrates that the algorithm has a strong capability in distinguishing complex background tissues (such as bone structures and imaging noise) from coronary vessels. The models were able to detect and localize coronary stenoses with high confidence levels (e.g., 53% and 75% as shown in Fig. 4). These visual results confirm that the proposed model is reliable not only in quantitative metrics but also in practical, real-world scenarios.

The findings of this study demonstrate that the use of lightweight object detection architectures such as EfficientDet-D3 and RetinaNet can provide an effective solution for bridging the gap between high accuracy and high processing speed in coronary artery stenosis detection. The achieved accuracy of 96.6% in the EfficientDet-D3 model confirms that, by leveraging mechanisms such as BiFPN, it is possible to deliver performance superior to that of more complex and computationally expensive models, even in noisy angiographic images. However, this study is subject to certain limitations: 1) although the dataset used is sufficiently large, it is single-center in nature, meaning that all images were acquired from a specific hospital using particular imaging devices. This may limit the generalizability of the model when applied to data from other clinical centers with different imaging protocols, and 2) the quality of input images remains a persistent challenge; the presence of interfering anatomical structures such as vertebrae or catheters in some frames may introduce noise and reduce detection accuracy. Furthermore, the current model is focused exclusively on single-vessel disease, and its performance in patients with multi-vessel involvement has not yet been fully evaluated. Future research is recommended to address these limitations by incorporating multi-center datasets and integrating advanced denoising techniques.