A dual-task deep learning framework for automated detection and classification of coronary artery lesions in invasive coronary angiography imaging.

objectiveThis study aims to develop and evaluate a dual-task deep learning framework for the simultaneous detection and classification of coronary lesions in invasive coronary angiography (ICA). MATERIALS AND

methodsA retrospective analysis was conducted using an annotated ICA dataset comprising 1234 patients (14,808 lesion-positive and 11,872 lesion-negative images), along with an external dataset of 135 cases. ICA video sequences were converted into 512 × 512 resolution images. A comprehensive preprocessing pipeline, including intensity normalization, noise suppression, and diverse data augmentation techniques (e.g., rotation, flipping, scaling, and brightness/contrast tuning), was applied to standardize image quality and enhance model generalizability. The proposed framework incorporates a detection module (utilizing YOLOv11, Swin Transformer, DETR, and Deformable DETR) and a classification module (featuring Vision Transformer (ViT), Swin Transformer, and ConvNeXt). The data were partitioned into training, validation, and testing subsets in an 80:10:10 ratio, and hyperparameter tuning was performed via grid search. Detection loss functions included Intersection over Union (IoU)-based losses, L1, and GIoU, while classification relied on weighted binary cross-entropy. Performance was evaluated using IoU, mAP, sensitivity, specificity, and AUC-PR.

resultsThe detection module exhibited consistent high performance across both internal and external datasets. Deformable DETR achieved superior results on the external dataset, with a mAP of 88.2%, IoU of 87.0%, sensitivity of 92.0%, and specificity of 90.1%, outperforming other models. For classification, the Swin Transformer reached an external accuracy of 92.8%, AUC-PR of 0.94, and a Cohen's Kappa of approximately 0.89, exceeding the performance of ConvNeXt and ViT. These results are comparable to human expert performance in coronary lesion detection, where accuracy ranges from 85 to 90%, demonstrating the system's robust diagnostic capability. The achieved sensitivity and specificity values are clinically meaningful, as they reduce the risk of both false positives and false negatives, crucial for ensuring timely and accurate treatment decisions in clinical practice.

conclusionThe proposed dual-task deep learning framework demonstrates a significant advancement in automated coronary lesion assessment by enabling accurate and efficient simultaneous detection and classification, thus supporting enhanced clinical decision-making.