Clinical implementation of 3D deep learning techniques in predicting touch-up lesions for atrial fibrillation patients undergoing cryoablation.
Authors
Affiliations (13)
Affiliations (13)
- Division of Cardiac Electrophysiology, Cardiovascular Center, Taipei Veterans General Hospital, Taipei, Taiwan.
- Institute of Clinical Medicine and Faculty of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan.
- Institute of Computer Science and Engineering, National Yang Ming Chiao Tung University, Hsinchu, Taiwan.
- Cardiovascular Center, Taichung Veterans General Hospital, Taichung, Taiwan.
- Institute of Biopharmaceutical Sciences, College of Pharmaceutical Sciences, National Yang Ming Chiao Tung University, Taipei, Taiwan.
- Institute of Biomedical Sciences, Academia Sinica, Taipei, Taiwan.
- Center for Fundamental Science, Kaohsiung Medical University, Kaohsiung, Taiwan.
- Biomedical Artificial Intelligence Academy, Kaohsiung Medical University, Kaohsiung, Taiwan.
- Institute of Statistics, National Yang Ming Chiao Tung University, Hsinchu, Taiwan.
- Department of Medical Research, Kaohsiung Medical University Hospital, Kaohsiung, Taiwan.
- Department of Statistics and Data Science, Cornell University, Ithaca, NY, USA.
- National Chung Hsing University, Taichung, Taiwan.
- Division of Cardiovascular Medicine, Department of Medicine, China Medical University Hospital, China Medical University, Taichung, Taiwan.
Abstract
Atrial fibrillation (AF) is a common heart rhythm disorder that can be treated with cryoballoon ablation (CBA). CBA occasionally requires additional radiofrequency-based touch-up ablation due to anatomical challenges. This study developed a 3D deep learning model to predict the complexity of CBA procedures and potentially reduce subsequent interventions, minimizing increased procedure times, costs, risks, and patient discomfort. We included 190 AF patients who underwent computed tomography (CT) scans at Taipei Veterans General Hospital from November 2014 to October 2020, divided into touch-up and non-touch-up groups. An 80:20 ratio was used to allocate patients to training and test sets, with an independent external validation set comprising 99 patients from October 2020 to August 2023. Three artificial intelligence (AI) models, PointNet, VoxNet, and an advanced version, CryoAI (VoxNet++), were developed to predict the need for touch-up ablation from 3D voxel-reconstructed CT images. CryoAI demonstrated the best overall discriminative performance among the tested models, achieving an area under the curve (AUC) of 84.07% in the internal test set, with a high positive predictive value (PPV) of 96.15%. In external validation, CryoAI maintained high performance with a PPV of 95.77%. Using a 60° curvature cutoff, all touch-up sites were localized to above-threshold regions in both the internal (n = 8) and external (n = 9) cohorts. Integrating Grad-CAM and a Gaussian Curvature module within our 3D Activation Visualization highlights critical zones for cryoballoon positioning and potential touch-up lesions, enhancing pre-procedural planning. The CryoAI model demonstrates promising discriminative ability for predicting the need for RF touch-up ablation in patients undergoing cryoballoon ablation for atrial fibrillation.