An Innovative Model for Diagnosing Lesions in Coronary Angiography Imagery Using an Improved YOLOv4 Model

Figures

• 

  Figure 1. Overall workflow diagram for intelligent diagnosis and prognostic assessment based on Coronary Angiography image analysis. The study workflow encompasses five main phases, from video data processing to model construction, pre-training, transfer learning, and final prognostic analysis. (A) Data Preprocessing. Key frames are extracted from coronary angiography videos, undergo desensitization, and are labeled for subsequent training. Colored Labeling Key: In the “labelling” step, colored boxes are used to mark different types of coronary lesions. Dark yellow presents limited meaningful lesions; Light green presents diffuse lesions; Red presents completely occlusive lesions. (B) Model. The architecture utilizes the improved YOLOv4 object detection algorithm. (C) Pre-training. The model is trained using the VOC 2007 dataset for training and validation, and the VOC 2008 dataset for testing, to acquire robust initial weights. (D) Transfer Learning: The pre-trained model is fine-tuned and tested on the clinical Coronary Angiography dataset. (E) Kaplan-Meier Analysis. Quantitatively assesses the correlation between algorithmically derived coronary artery lesion characteristics and the…

• 

  Figure 2. Flowchart of coronary angiography image inclusion and algorithm application.

• 

  Figure 3. Diagram of the improved YOLOv4 model. (A) CSPDarknet53 Backbone. The feature extraction network that processes the Inputs (512, 512, 3) image and yields features at various scales. (B) Modified Spatial Pyramid Pooling (SPP) Module. The pooling layer sizes are optimized to 1, 2, 4, 8, and 16, enhancing the model’s multiscale feature aggregation capacity for accurate localization of variably sized lesions. (C) Path Aggregation Network (PANet). The feature aggregation modules that fuse high-level and low-level features through convolution, upsampling, downsampling, and concatenation to enrich the feature map. (D) YOLO Detection Heads. The standard detection heads that output final predictions (bounding boxes and classification) at three different scales (e.g., (16, 16, 255), (32, 32, 255), and (64, 64, 255)).

• 

  Figure 4. Coronary angiography image desensitization and labelling. (A1–A3). Original images of coronary angiography from patients. (B1–B3). Desensitization of relative original images. (C1–C3). Labelling of the narrowed coronary arteries image, dark yellow in (C1) presents limited meaningful lesion, light green in (C2) presents diffuse lesion, red in (C3) presents completely occlusive lesions.

• 

  Figure 5. Schematic diagram of SPP module before and after improvement.

• 

  Figure 6. Line graph showing validation of improved YOLOv4 model on the VOC 2008 dataset. The VOC 2008 dataset was randomly divided into six equal parts for testing the pre-trained model. The improved model’s training batch was set to 100, with a batch size of 4. The initial learning rate was set to 0.01, with a minimum value of 0.0001. Average precision (AP) values for each object category were computed.

• 

  Figure 7. Comparison of the improved YOLOv4 model for the identification of different lesions in coronary angiography images. The red box represents the true value of manual labeling; the green box represents the predicted value of the lesion by the improved YOLOv4 model.

• 

  Figure 8. Association between coronary lesions and MACE occurrence in AMI patients. (A) Association between meaningful stenosis lesions and MACEs. (B) Association between diffuse stenosis lesions and MACEs. (C) Association between complete occlusive lesions and MACEs. (D) Association between combined lesions and MACEs. AMI: acute myocardial infarction, MACEs: major adverse cardiovascular events. The shaded area around the curve represents the 95% confidence interval.