Optimizing Input Selection for Cardiac Model Training and Inference: An Efficient 3D CNN-based Approach to Automate Coronary Angiogram Video Selection

Abstract Background Research leveraging deep learning (DL) for medical image analysis is increasingly using dynamic coronary angiography from cardiac catheterizations to train neural networks. Yet, an efficient, automatic method to select appropriate dynamic images for training is still largely missing. Methods We developed DL models using 254 coronary angiographic studies from the Mayo Clinic. We utilized two state-of-the-art Convolutional Neural Networks (CNN: ResNet and X3D), to identify low quality angiograms through binary classification (satisfactory/unsatisfactory). Ground truth for the quality of the input angiogram was determined by two experienced cardiologists. We validated the developed model in an independent dataset of 3,208 procedures from 3 Mayo sites. Results 3D-CNN models outperformed their 2D counterparts, with the X3D model achieving superior performance across all metrics (AUC 0.98, precision 0.86, and sensitivity 0.89). The 2D models processed the video clips faster than 3D models. Despite having a 3D architecture, the X3D model had lower computational demand (2.56 GMAC) and parameter count (2.98 M) than 2D models. When validating models on the independent dataset, slight decreases in AUC and sensitivity were observed but accuracy and specificity remained robust (0.88 and 0.89, respectively for the X3D model). Conclusion We developed a rapid and effective method for automating the selection of coronary angiogram video clips using 3D-CNNs, potentially improving model accuracy and efficiency in clinical applications. The X3D-S model demonstrates a balanced trade-off between computational efficiency and complexity, making it suitable for real-life clinical applications.