Detection and frequency investigation of common coronary artery vari̇ati̇ons i̇n cardi̇ac CT images using the yolov-10 archi̇tecture based on artificial intelli̇gence

Aim Advances in imaging technology in recent years have enabled more detailed examination of the anatomical structure of coronary arteries. Coronary Angiography Computed Tomography (CACT) is frequently preferred as a non-invasive method for evaluating coronary artery anatomy and variations. The primary objective of this study is to develop a deep learning-based system capable of detecting common coronary artery variations in CTA images. The proposed system aims to provide a clinically applicable, reliable, and effective solution by leveraging the balance of speed and accuracy offered by the YOLOv10 architecture. Materials and Methods In this study, the deep learning-based YOLOv10 algorithm was used to detect coronary artery variations. In the study, a dataset consisting of 2,837 multi-slice cardiac CT images from 20 patients selected from 100 patients who presented to a tertiary hospital with cardiac problems after exclusion criteria were applied. The images were divided into three subgroups: 2,263 for training, 287 for validation, and 287 for testing. The dataset was distributed in a balanced manner to evaluate the model's accuracy and effectiveness in different scenarios. Results The YOLOv10 model has made highly accurate predictions for coronary artery variations and the normal class. The model's classification capacity has been found to be reliable, with the vast majority of predictions classified correctly. In addition, the detected bounding boxes have provided a strong foundation for identifying coronary artery variations and demonstrated YOLOv10's high potential in medical image analysis. In coronary artery variation detection, the balance between precision, recall, and mAP is important depending on the application's requirements. In cases requiring high accuracy, YOLOv10-X stands out, while YOLOv10-L is preferred for more balanced and generalizable results.