Clinical implementation of 3D deep learning techniques in predicting touch-up lesions for atrial fibrillation patients undergoing cryoablation - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Int J Cardiol Heart Vasc . 2026 Apr 3;64:101912. doi: 10.1016/j.ijcha.2026.101912 Search in PMC Search in PubMed View in NLM Catalog Add to search Clinical implementation of 3D deep learning techniques in predicting touch-up lesions for atrial fibrillation patients undergoing cryoablation Chih-Min Liu Chih-Min Liu a Division of Cardiac Electrophysiology, Cardiovascular Center, Taipei Veterans General Hospital, Taipei, Taiwan b Institute of Clinical Medicine and Faculty of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan Find articles by Chih-Min Liu a, b, 1 , Wei-Wen Chen Wei-Wen Chen c Institute of Computer Science and Engineering, National Yang Ming Chiao Tung University, Hsinchu, Taiwan Find articles by Wei-Wen Chen c, 1 , Shih-Lin Chang Shih-Lin Chang a Division of Cardiac Electrophysiology, Cardiovascular Center, Taipei Veterans General Hospital, Taipei, Taiwan b Institute of Clinical Medicine and Faculty of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan Find articles by Shih-Lin Chang a, b, ⁎ , Chien-Chao Tseng Chien-Chao Tseng c Institute of Computer Science and Engineering, National Yang Ming Chiao Tung University, Hsinchu, Taiwan Find articles by Chien-Chao Tseng c , Ching-Chun Huang Ching-Chun Huang c Institute of Computer Science and Engineering, National Yang Ming Chiao Tung University, Hsinchu, Taiwan Find articles by Ching-Chun Huang c , Yenn-Jiang Lin Yenn-Jiang Lin a Division of Cardiac Electrophysiology, Cardiovascular Center, Taipei Veterans General Hospital, Taipei, Taiwan b Institute of Clinical Medicine and Faculty of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan k Cardiovascular Center, Taichung Veterans General Hospital, Taichung, Taiwan Find articles by Yenn-Jiang Lin a, b, k , Li-Wei Lo Li-Wei Lo a Division of Cardiac Electrophysiology, Cardiovascular Center, Taipei Veterans General Hospital, Taipei, Taiwan b Institute of Clinical Medicine and Faculty of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan Find articles by Li-Wei Lo a, b , Yu-Feng Hu Yu-Feng Hu a Division of Cardiac Electrophysiology, Cardiovascular Center, Taipei Veterans General Hospital, Taipei, Taiwan b Institute of Clinical Medicine and Faculty of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan d Institute of Biopharmaceutical Sciences, College of Pharmaceutical Sciences, National Yang Ming Chiao Tung University, Taipei, Taiwan e Institute of Biomedical Sciences, Academia Sinica, Taipei, Taiwan Find articles by Yu-Feng Hu a, b, d, e , Fa-Po Chung Fa-Po Chung a Division of Cardiac Electrophysiology, Cardiovascular Center, Taipei Veterans General Hospital, Taipei, Taiwan b Institute of Clinical Medicine and Faculty of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan Find articles by Fa-Po Chung a, b , Ting-Yung Chang Ting-Yung Chang a Division of Cardiac Electrophysiology, Cardiovascular Center, Taipei Veterans General Hospital, Taipei, Taiwan b Institute of Clinical Medicine and Faculty of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan Find articles by Ting-Yung Chang a, b , Chin-Yu Lin Chin-Yu Lin a Division of Cardiac Electrophysiology, Cardiovascular Center, Taipei Veterans General Hospital, Taipei, Taiwan b Institute of Clinical Medicine and Faculty of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan Find articles by Chin-Yu Lin a, b , Tze-Fan Chao Tze-Fan Chao a Division of Cardiac Electrophysiology, Cardiovascular Center, Taipei Veterans General Hospital, Taipei, Taiwan b Institute of Clinical Medicine and Faculty of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan Find articles by Tze-Fan Chao a, b , Ta-Chuan Tuan Ta-Chuan Tuan a Division of Cardiac Electrophysiology, Cardiovascular Center, Taipei Veterans General Hospital, Taipei, Taiwan b Institute of Clinical Medicine and Faculty of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan Find articles by Ta-Chuan Tuan a, b , Jo-Nan Liao Jo-Nan Liao a Division of Cardiac Electrophysiology, Cardiovascular Center, Taipei Veterans General Hospital, Taipei, Taiwan b Institute of Clinical Medicine and Faculty of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan Find articles by Jo-Nan Liao a, b , Ling Kuo Ling Kuo a Division of Cardiac Electrophysiology, Cardiovascular Center, Taipei Veterans General Hospital, Taipei, Taiwan b Institute of Clinical Medicine and Faculty of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan Find articles by Ling Kuo a, b , Cheng-I Wu Cheng-I Wu a Division of Cardiac Electrophysiology, Cardiovascular Center, Taipei Veterans General Hospital, Taipei, Taiwan b Institute of Clinical Medicine and Faculty of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan Find articles by Cheng-I Wu a, b , Shin-Huei Liu Shin-Huei Liu a Division of Cardiac Electrophysiology, Cardiovascular Center, Taipei Veterans General Hospital, Taipei, Taiwan b Institute of Clinical Medicine and Faculty of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan Find articles by Shin-Huei Liu a, b , Jacky Chung-Hao Wu Jacky Chung-Hao Wu f Center for Fundamental Science, Kaohsiung Medical University, Kaohsiung, Taiwan g Biomedical Artificial Intelligence Academy, Kaohsiung Medical University, Kaohsiung, Taiwan h Institute of Statistics, National Yang Ming Chiao Tung University, Hsinchu, Taiwan Find articles by Jacky Chung-Hao Wu f, g, h , Henry Horng-Shing Lu Henry Horng-Shing Lu g Biomedical Artificial Intelligence Academy, Kaohsiung Medical University, Kaohsiung, Taiwan h Institute of Statistics, National Yang Ming Chiao Tung University, Hsinchu, Taiwan i Department of Medical Research, Kaohsiung Medical University Hospital, Kaohsiung, Taiwan j Department of Statistics and Data Science, Cornell University, Ithaca, NY, USA Find articles by Henry Horng-Shing Lu g, h, i, j, ⁎ , Shih-Ann Chen Shih-Ann Chen a Division of Cardiac Electrophysiology, Cardiovascular Center, Taipei Veterans General Hospital, Taipei, Taiwan b Institute of Clinical Medicine and Faculty of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan k Cardiovascular Center, Taichung Veterans General Hospital, Taichung, Taiwan l National Chung Hsing University, Taichung, Taiwan m Division of Cardiovascular Medicine, Department of Medicine, China Medical University Hospital, China Medical University, Taichung, Taiwan Find articles by Shih-Ann Chen a, b, k, l, m Author information Article notes Copyright and License information a Division of Cardiac Electrophysiology, Cardiovascular Center, Taipei Veterans General Hospital, Taipei, Taiwan b Institute of Clinical Medicine and Faculty of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan c Institute of Computer Science and Engineering, National Yang Ming Chiao Tung University, Hsinchu, Taiwan d Institute of Biopharmaceutical Sciences, College of Pharmaceutical Sciences, National Yang Ming Chiao Tung University, Taipei, Taiwan e Institute of Biomedical Sciences, Academia Sinica, Taipei, Taiwan f Center for Fundamental Science, Kaohsiung Medical University, Kaohsiung, Taiwan g Biomedical Artificial Intelligence Academy, Kaohsiung Medical University, Kaohsiung, Taiwan h Institute of Statistics, National Yang Ming Chiao Tung University, Hsinchu, Taiwan i Department of Medical Research, Kaohsiung Medical University Hospital, Kaohsiung, Taiwan j Department of Statistics and Data Science, Cornell University, Ithaca, NY, USA k Cardiovascular Center, Taichung Veterans General Hospital, Taichung, Taiwan l National Chung Hsing University, Taichung, Taiwan m Division of Cardiovascular Medicine, Department of Medicine, China Medical University Hospital, China Medical University, Taichung, Taiwan ⁎ Corresponding authors at: Division of Cardiac Electrophysiology, Cardiovascular Center, Taipei Veterans General Hospital, No. 201, Sec. 2, Shih-Pai Road, Beitou Dist., Taipei City 112, Taiwan (S.-L. Chang). Biomedical Artificial Intelligence Academy, Kaohsiung Medical University, No. 100, Shih-Chuan 1st Road, Kaohsiung 80708, Taiwan (H. H.-S. Lu). [email protected] [email protected] 1 Equal contribution: Chih-Min Liu and Wei-Wen Chen. Received 2025 Nov 26; Revised 2026 Mar 9; Accepted 2026 Mar 22; Collection date 2026 Jun. © 2026 Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). PMC Copyright notice PMCID: PMC13087757 PMID: 42005401 Highlights • Developed a 3D deep learning model (CryoAI) to predict touch-up need in cryoablation. • CryoAI (VoxNet++) outperformed PointNet and VoxNet in both internal and external datasets. • Grad-CAM-based 3D visualization provided an interpretable model that focuses on atrial structures. • Gaussian Curvature analysis identified high-risk regions for incomplete ablation contact. • CryoAI enables preoperative planning to possibly reduce procedural time and improve outcomes. Keywords: Atrial fibrillation, Cryoballoon ablation, Touch-up ablation, Artificial intelligence, 3D activation visualization, 3D Gaussian curvature Abstract Background 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. Methods 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. Results 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. Conclusions The CryoAI model demonstrates promising discriminative ability for predicting the need for RF touch-up ablation in patients undergoing cryoballoon ablation for atrial fibrillation. 1. Introduction Atrial fibrillation (AF) is the most common cardiac rhythm disorder characterized by irregular and often rapid heartbeats [1] . It is associated with increased morbidity, including conditions such as stroke and heart failure, and mortality, significantly burdening the healthcare system and leading to a decreased quality of life [2] . Catheter ablation has emerged as a standard treatment option for patients with AF who do not respond well to medications [3] , [4] . Catheter ablation has been demonstrated to be more effective than antiarrhythmic drug therapy in maintaining sinus rhythm [5] . The safety profile of cryoballoon ablation (CBA) has been established for patients who do not respond to drug therapy, with serious procedure-related adverse events being rare [6] , [7] . This approach uses a special catheter with a balloon at its tip, inflated and cooled with extreme cold to create surrounding scars over the pulmonary veins (PVs) [8] . CBA offers several advantages, including shorter procedure times and potentially less collateral tissue damage, but its effectiveness may vary according to anatomy [9] , [10] , [11] . Some cases present anatomical challenges that can affect the effectiveness of CBA, potentially leading to the need for additional radiofrequency (RF)-based “touch-up” ablation procedures [12] . The additional touch-up ablation not only prolongs procedural times and costs but also worsens ablation outcomes [13] , [14] . Anatomical factors such as the size of the PVs and the angles between the PVs and the left atrium (LA) are crucial for successful CBA without the need for additional touch-up ablations [15] . However, to date, no studies have utilized artificial intelligence (AI) to analyze the spatial location and morphology of the LA and PVs. A voxel, composed of vertices, edges, and faces, defines the shape of polyhedral objects and can represent spatial locations with minimal information [16] . We hypothesize that a three-dimensional (3D) deep learning model trained on voxel representation datasets can accurately predict the need for additional touch-up ablations during the CBA procedure. A 3D deep learning model was developed, capable of learning from voxel data and classifying whether CBA requires touch-up ablations. By establishing a novel evaluation method for CBA, this model might enable preoperative assessments to determine whether patients will complete treatment or require additional touch-up ablation. 2. Methods 2.1. Study population The study included patients who underwent pulmonary vein computed tomography (PVCT) scans at Taipei Veterans General Hospital between November 2014 and October 2020. The cohort consisted of 190 patients with AF undergoing CBA (Arctic Front cryoballoon, Medtronic, Inc., Minneapolis, MN, USA, or PolarX, Boston Scientific, Marlborough, MA, USA). CBA was performed using either a conventional application strategy or a time-to-pulmonary vein isolation (TT-PVI)-guided dosing protocol [17] , at the discretion of the attending electrophysiologist, in accordance with contemporary clinical practice and prior studies. In the conventional approach, two to three cryoapplications were typically delivered to each PV, with individual application durations ranging from 120 to 240 s, guided by real-time procedural parameters including PV occlusion quality, balloon nadir temperature, and thaw time. In the TT-PVI-guided dosing protocol, one or two cryoapplications were delivered according to predefined TT-PVI algorithms, with balloon repositioning performed when effective TT-PVI could not be achieved. PV isolation was systematically assessed after CBA using a circular mapping catheter to confirm both entrance and exit block. Touch-up ablation was performed only when residual PV conduction persisted despite repeated and adequately positioned cryoballoon applications, according to the institutional workflow and operator decision thresholds applied at our center. Those patients were categorized based on procedural reports into two groups: those requiring additional touch-up ablation and those not requiring touch-up ablation. The patients who required additional RF touch-up ablation over PVs following cryoablation were defined as the touch-up group. In contrast, the non-touch-up group did not require additional RF touch-up ablation. Touch-up ablation was restricted to sealing residual PV conduction gaps; non-PV triggers, linear lesions, and other substrate modifications were not included. PV gaps were treated using an open-irrigated tip catheter (ThermoCool, Biosense Webster; FlexAbility, St. Jude Medical/Abbott; INTELLANAV, Boston Scientific) or a non-irrigated tip catheter (4 mm, St. Jude or Biosense Webster), guided by a 3D electroanatomic mapping system, according to the physicians’ preference. For the open-irrigated tip catheter, RF power was 25–35 W with a target duration of 40 s per lesion, and temperature was capped at ≤40 °C. For the non-irrigated tip catheter, RF power was 50 W with a target duration of 40 s per lesion, and the temperature was set at 50°C. We randomized 80% of the patients to the training set (151 patients) and 20% to the test set (39 patients), with the ratio of touch-up to non-touch-up groups being approximately 1:4. To address class imbalance, all data splits were generated using stratified sampling, and synthetic oversampling techniques were not applied due to the high dimensionality of 3D voxel data. Instead, class imbalance was handled using a class-weighted binary cross-entropy loss during training, with performance evaluated using multiple complementary metrics to avoid bias toward positive predictive value. The study flowchart is presented in Fig. 1 . Furthermore, an additional independent test set comprising 99 patients treated between October 2020 and August 2023 was obtained for external validation. This dataset included nine patients requiring a touch-up and 90 patients who did not. Fig. 1. Open in a new tab Study flowchart with temporal split. A temporal cutoff at October 2020 (vertical tick) divided the cohort into Cohort A (patients enrolled before October 2020) and Cohort B (patients enrolled from October 2020 onward). Cohort A was randomly partitioned into training and test sets for model development, whereas Cohort B was held out for external validation. Medical histories, laboratory blood tests, and medications were collected prior to catheter ablation. This information was sourced from the medical records of primary and secondary referral hospitals, outpatient and emergency visits, the Collaboration Center of Health Information Application (CCHIA), and the Ministry of Health and Welfare in Taiwan. The Ninth and Tenth Revisions of the International Classification of Diseases (ICD-9 and ICD-10) codes were employed to ascertain the presence of medical conditions. These details have been outlined in the previous studies [18] , [19] . 2.2. 3D deep learning model In this study, voxel-reconstructed CT images referred to the original 3D CT volumes produced by the scanner after image reconstruction, in which cardiac structures were represented as stacks of isotropic volumetric pixels (voxels). These voxel-based CT volumes constituted the primary imaging data used for manual or semi-automated segmentation of the LA and PVs. Following segmentation, the anatomical structures were converted into derivative geometric representations for subsequent analysis. Specifically, the segmented left atrial–pulmonary vein anatomy was first transformed into a 3D surface mesh (triangulated surface model), which enabled quantitative assessment of surface geometry, including curvature-based analysis. This surface mesh was then resampled into a fixed-size 3D voxel grid representing binary anatomical occupancy, thereby standardizing anatomical scale and orientation across patients. These derived voxelized anatomical representations served as the direct input to the VoxNet-based deep learning model for the prediction of touch-up ablation sites. Volume data of the PVs and LA from PVCT images in DICOM format were converted to obj file format. The proposed method comprises two stages, as illustrated in Fig. 2 . In the initial stage, the patient's PVCT images are reconstructed into 3D mesh and voxel images. The initial step involves extracting data from the XML markup files generated by the imaging equipment, which is then converted into 3D image objects. This transformation process enables the capture of diverse features, including 3D spatial relationships, angular variations, and curvature. Three functional procedures are defined to obtain vectors, normal vectors, and polygon data. Subsequently, the Trimesh library is employed to generate 3D image objects in both mesh and voxel formats. In the second stage of the study, a 3D deep learning model was constructed to determine whether additional touch-up ablation is necessary. Fig. 2. Open in a new tab Two-stage process for transforming pulmonary vein and left atrium structural data from PVCT images into 3D models for deep learning analysis. The initial stage involves converting volume data from the DICOM format of the PVCT images into an obj file format. This transformation starts with the extraction of data from XML markup files produced by the imaging equipment, which is subsequently converted into 3D image objects. This process allows for capturing complex features such as 3D spatial relationships, angular variations, and curvature. Three functional procedures are defined for obtaining vectors, normal vectors, and polygon data, utilizing the Trimesh library to generate 3D image objects in both mesh and voxel formats. In the second stage, a 3D deep learning model (depicted as a blue cube) is created to evaluate whether additional touch-up ablation is required. This approach demonstrates the detailed processing of medical imaging data through advanced 3D reconstruction and deep learning techniques. Abbreviation: PVCT, pulmonary vein computed tomography; 3D, three-dimensional. In the final phase of the study, three AI models were developed: PointNet, VoxNet, and our proposed enhanced version of VoxNet (VoxNet++), named CryoAI. The 3D mesh images created in the initial phase are provided as point clouds to the PointNet deep learning model and as voxel grids to the VoxNet and CryoAI deep learning models. The predictive performance of the three models is evaluated, and the results are presented in 3D visualizations. 2.3. 3D Gaussian curvature computation After PV segmentation from PVCT, the left atrial-PV surface was reconstructed as a triangulated 3D mesh. Local surface geometry was quantified using Gaussian curvature, defined at each mesh vertex as the angular defect (∑α i subtracted from 2π) within a geodesic neighborhood (radius = 1 mm). Curvature values were expressed in degrees to facilitate anatomical interpretability. For visualization and analysis, curvature maps were generated at representative thresholds (30°, 60°, and 90°) to model increasing degrees of local convexity or tapering that may affect cryoballoon contact ( Fig. S1 ). These curvature maps were overlaid on the 3D PV anatomy and subsequently integrated into the voxel-based pipeline for downstream explainable analysis. We computed a discrete approximation of Gaussian Curvature on the left atrial surface mesh reconstructed from PVCT ( Fig. S2 ) [20] . For each mesh vertex v with incident faces { f i} [15] , [21] , we first computed the vertex defect δ v = 2 π - i ∑ α i , where α i is the corner angle at v in face f i. For every surface point p , neighboring vertices within an Euclidean radius r were identified using a k -dimensional (KD) tree, and face angles around each neighbor were calculated. The curvature value at p was then obtained by summing the vertex defects of all vertices in this neighborhood: Kr p = v ∈ Br p ∑ δ v . Optionally, we normalized by the local surface area Ar(p) (estimated from one-third of the areas of incident triangles of vertices in Br(p) ) to yield an area-normalized estimate K^r(p) = Kr(p)/Ar(p) . Curvature values were mapped to a color scale; regions exceeding a user-defined threshold were highlighted in red for visualization. Unless stated otherwise, we used r = 1 mm for LA visualizations. In other words, the output is the discrete Gaussian Curvature obtained by summing the vertex defects at all vertices within the radius for each point. 2.4. PointNet model PointNet is a deep-learning point cloud classification and segmentation model [22] . A point cloud can be represented as an N × 3 matrix, where N denotes the number of points in space. PointNet uses a T-Net to ensure the invariance of the point cloud in space. It then applies a series of MLPs (Multi-Layer Perceptrons), implemented as 1 × 1 convolutions, to generate an N × 1024 feature map. A max pooling operation is then used to produce a 1024-dimensional global feature, which addresses the order invariance of the point cloud. Finally, several linear layers are added to perform the classification. While PointNet excels in extracting global features, its ability to capture local details is relatively weak, which may limit its performance in tasks requiring fine-grained analysis. The model's complexity can also lead to higher computational resource requirements for training and inference. 2.5. VoxNet model The VoxNet model, comprising a straightforward yet productive convolutional neural network (CNN), accepts voxel grids as input [23] , [24] . Using voxel grids to represent 3D data offers several advantages over traditional 3D data representations, such as point clouds. A significant benefit is that voxel grids obviate the necessity for extensive preprocessing, including feature extraction, matching, and fusion, which can increase the computational load and latency. VoxNet employs a voxel representation of 3D point cloud data, with the voxel size set to 0.1 m. This conversion reduces the original data from millions of points to hundreds of thousands of voxels, effectively decreasing the computational complexity and improving the efficiency of the process. Conventional CNNs perform convolution operations across the entire input data, resulting in high computational costs. In contrast, VoxNet employs local pooling operations, which perform convolutions only on local regions within the voxel grid. This approach enhances computational efficiency. While VoxNet gradually extracts features using convolutional kernels of varying sizes, its single-path sequential model structure may limit its ability to handle complex data. We proposed the VoxNet++ (CryoAI) model, which features a dual-path structure design to address this limitation. This design enhances the model's capacity to capture comprehensive data characteristics and improves its generalization ability. 2.6. CryoAI (VoxNet++) model CryoAI was implemented in TensorFlow/Keras as a VoxNet-based 3D CNN operating on 121 × 121 × 121 pulmonary vein CT voxel grids and trained as a binary classifier (touch-up vs. non–touch-up). Model training used a batch size of 32 with class-weighted binary cross-entropy loss, and hyperparameters were optimized using a Hyperband search (KerasTuner) with validation AUC as the objective, including learning rate (10⁻ 4 –10⁻⁶), L2 regularization strength (10⁻⁶–10⁻ 2 ), dropout ratio, and optimizer choice. The final model employed an Adam-based optimizer with an initial learning rate of 1 × 10⁻ 4 and L2 regularization on fully connected layers. Model performance was estimated using 4-fold stratified cross-validation, with training up to 400 epochs per fold and early stopping (patience = 5) together with a Reduce-on-Plateau learning rate scheduler. In Fig. S3 , we illustrated the network architecture of VoxNet++ (CryoAI) and VoxNet. Both models were designed to process 3D data, such as medical imaging. They were based on 3D CNNs for feature extraction. 3D CNNs could effectively extract local and global features from 3D data, eliminating the need for extensive manual preprocessing, such as feature extraction. The VoxNet++ (CryoAI) method represented a significant advancement over VoxNet. The distinguishing characteristics were the dual-path design structure. Although VoxNet demonstrated proficiency in feature extraction, its sequential model structure constrained its capacity to extract and analyze multi-scale features simultaneously. It employed a sizeable convolutional kernel (5 × 5 × 5, stride 2) to extract broad features, followed by a small convolutional kernel (3 × 3 × 3, stride 1) for fine-tuning. In contrast, VoxNet++ (CryoAI) overcame this limitation with a more flexible and adaptable functional architecture, rendering it suitable for more complex topologies, such as those with multiple inputs and layer sharing. One of the key innovations in VoxNet++ (CryoAI) was the introduction of a pose classifier, which enabled the learning of spatial position and orientation features from voxel data. The dual-path design was a crucial aspect of VoxNet++ (CryoAI), as it allowed the model to process input data through two independent paths: one focusing on the coarse pose and the other on the fine pose. This unique design empowered VoxNet++ (CryoAI) to integrate features of different scales, enhancing its ability to recognize objects at various scales. By capturing features at different levels simultaneously, VoxNet++ (CryoAI) could extract multi-scale features from the same input data, thereby improving its ability to capture comprehensive data characteristics and enhancing generalization. We manually added L2 regularization to all dense layers to help prevent overfitting. In summary, the advantages of the VoxNet++ model were as follows: • Flexibility: The dual-path design allowed the model to handle features of different scales more flexibly. • Generalization: Combining different feature extraction methods improved the generalization ability. 3. Results The distribution between the two groups (touch-up and non-touch-up) was about 1:4, with 37 patients in the touch-up group and 153 patients in the non-touch-up group. The baseline characteristics of the study population are shown in Table S1 . The proposed network architecture's efficacy was assessed by utilizing a range of performance metrics, including accuracy, recall, precision, F1 score, and area under the curve (AUC). A comparison of the performance of the various models was conducted using PointNet, VoxNet, and the enhanced version of VoxNet, named VoxNet++ (CryoAI). Table 1 presents the performance comparison results on the internal test set and on an additional independent test set. Table 1. Comparison of model performance metrics on the internal and external test sets. Internal test set Input data Model Accuracy Sensitivity Specificity Precision NPV F1 score AUC Point cloud PointNet 66.67 67.74 62.50 87.50 33.33 76.36 65.12 3D Voxel grid VoxNet 84.62 87.10 75.00 93.10 60.00 90.00 81.05 3D Voxel grid VoxNet++ (CryoAI) 82.05 87.50 80.65 96.15 53.85 87.72 84.07 External test set Input data Model Accuracy Sensitivity Specificity Precision NPV F1 score AUC Point cloud PointNet 36.00 30.77 88.89 96.55 11.27 46.67 59.44 3D Voxel grid VoxNet 81.00 83.52 55.56 95.00 25.00 88.89 68.89 3D Voxel grid VoxNet++ (CryoAI) 74.00 74.73 66.67 95.77 20.69 83.95 70.56 Open in a new tab The numbers are presented with percentages (%). Precision stands for positive predictive value (PPV). Abbreviation: AUC, area under the curve; NPV, negative predictive value; 3D, three-dimensional. 3.1. Evaluation of performance measures Table 1 illustrates that our proposed VoxNet++ (CryoAI) model architecture demonstrated superior performance in terms of AUC, achieving a score of 84.07% and a positive predictive value (PPV) of 96.15% compared to the other models in the internal test set. This outcome suggested that the dual-path design was more effective in capturing global and fine-grained local features, making it a more suitable approach for complex 3D topologies. Upon testing the model on an independent dataset collected from October 2020 to August 2023 (comprising 99 patients, 9 in the touch-up group and 90 in the non-touch-up group), the AUC and PPV were found to be 70.56% and 95.77%, respectively, as shown in Table 1 . This performance was also observed to be higher than that of VoxNet (68.89%) and PointNet (59.44%). Using a predefined CryoAI probability cutoff of 0.5, confusion-matrix analyses were performed to provide transparency for the reported performance metrics. In the internal test cohort, CryoAI identified 7 true positives, 25 true negatives, 6 false positives, and 1 false negative, corresponding to a sensitivity of 87.50% and a specificity of 80.65%. In the temporally independent external validation cohort, CryoAI yielded 6 true positives, 67 true negatives, 23 false positives, and 3 false negatives. These confusion-matrix counts are shown in Fig. S4 and directly support the reported discrimination metrics for both cohorts. Subgroup analyses were performed to explore whether clinical variables and PV anatomical features were associated with the need for touch-up ablation. With respect to PV anatomy (shown in Table S2 ), the prevalence of right middle PVs (touch-up vs. non-touch-up: 32.4% vs. 30.7%, P = 0.841) and left common PVs (touch-up vs. non-touch-up: 10.8% vs. 12.4%, P = 0.789) did not differ significantly between groups. These findings indicate that conventional categorical PV anatomical variants alone in our cohort were not significantly associated with touch-up ablation. In contrast, CryoAI predictions were derived from comprehensive 3D geometric information, suggesting that higher-order anatomical features beyond standard PV classifications may underlie the model’s predictive capability. 3.2. 3D Activation visualization Activation Visualization techniques, such as Grad-CAM (Gradient-weighted Class Activation Mapping) [25] , offer immense potential when applied to clinical contexts in AI. These techniques enable AI systems to interpret critical data features, providing diagnosticians with insights from various spatial perspectives to observe subtle changes. They also aid in understanding the decision-making processes of neural networks, revealing which features in the input data are crucial for the model's predictions. Our 3D Activation Visualization tool distinguished itself from traditional 2D visualization methods by employing spatial transformation to map Grad-CAM heatmaps to 3D voxel positions. This approach only displayed heatmaps (activation values multiplied by weighted vectors) within the annotated object boundaries. Fig. 3 demonstrates an example of the Grad-CAM visualization of a 3D LA voxel image, excluding areas outside the heart's boundaries. Additionally, our tool offered functionalities such as rotation, zoom-in, and zoom-out, allowing diagnosticians to observe fine details from any angle of interest easily, as the AI focuses on. Fig. 3. Open in a new tab 3D Activation Visualization of an LA voxel image using Grad-CAM. The heatmap highlighted regions within the heart's boundaries where the AI model focuses its attention. The tool allowed for rotation, zoom-in, and zoom-out functionalities, enabling diagnosticians to examine fine details from various angles. Areas outside the heart were excluded from the visualization, providing a clear and focused view of the critical features influencing the model's predictions. The deeper colors indicate the regions where the model focuses more during classification. The transition from deeper to lighter colors represents the degree of importance from high to low (red → purple → yellow). The values 0 and 1 are used to determine the object’s boundary, where 1 indicates the inside of the object and 0 indicates the outside. The Activation Visualization results are displayed only for the regions marked with 1. Abbreviation: AI, artificial intelligence; Grad-CAM, Gradient-weighted Class Activation Mapping; LA, left atrium; 3D, three-dimensional. 3.3. 3D Gaussian curvature Before performing CBA in patients with AF, it would be beneficial for electrophysiologists to have a tool that helps evaluate whether inserting a cryoballoon on a given surface is feasible. This evaluation can be enhanced by calculating the curvature of the surface to identify the optimal surgical site. For example, a region with a high degree of curvature may present greater challenges during surgical intervention. The greater the curvature, the more the surface bends, which can impede the ability of the cryoballoon to adhere correctly. Conversely, a region with low curvature is flatter and more accessible for placing the cryoballoon. Therefore, the larger the curvature, the harder it is to position the cryoballoon; the smaller the curvature, the easier it is. We developed a tool that can instantly calculate the Gaussian Curvature of the surface of a 3D voxel object. It computes the local curvature value based on a specified radius. The Gaussian Curvature calculation is as follows: first, define vertex defects as 2π minus the sum of the angles of every face that includes that vertex. Given inputs (mesh, points, radius), where points represent points in space, the tool uses a KD tree to find the nearest points within the specified radius and then calculates the face angles formed by these points. The output is the discrete Gaussian Curvature obtained by summing the vertex defects at all vertices within the radius for each point. Fig. S5 shows an example of 3D LA curvature with a local radius range of 1 mm. Areas in red indicate regions where the curvature exceeds the threshold. Electrophysiologists are particularly interested in the curvature around the PV region, as this is typically where the CBA is performed. We observed that patients who required additional touch-up ablation had higher curvature values in these regions, whereas those who did not require further touch-up had lower curvature values. For example ( Fig. S1 ), regions with low Gaussian curvature (∼30°) corresponded to smoothly contoured PV ostia, where the cryoballoon geometry demonstrated favorable surface conformity. Intermediate curvature (∼60°) identified transitional regions with reduced balloon–tissue contact, while high curvature (∼90°) consistently localized sharply tapered or conical PV segments where cryoballoon fitting was suboptimal. Importantly, a curvature threshold of 60° successfully captured all clinically observed touch-up ablation sites, yielding 100% sensitivity with a specificity of 62.5%. Every touch-up site fell within a region exceeding this value in both the internal (n = 8) and external (n = 9) cohorts. Although the specificity was modest, this threshold was intentionally selected to prioritize sensitivity, ensuring that no potential touch-up sites were missed. Empirically, lower curvature thresholds increased false-positive regions without improving sensitivity, whereas higher thresholds reduced sensitivity by failing to capture all confirmed touch-up locations. Given that the Gaussian curvature analysis was designed as an explainable visualization tool rather than a standalone diagnostic classifier, the 60° cutoff provides a practical balance by highlighting all anatomically high-risk regions for further assessment. These findings indicate that elevated Gaussian curvature reflects anatomically constrained regions prone to incomplete cryoballoon lesion formation and provides a quantitative, anatomically interpretable marker linking PV geometry to the need for touch-up ablation. This approach also enables comprehensive identification of potential touch-up areas and serves as a preprocedural warning aid for electrophysiologists. 4. Discussion This study investigated the potential of deep learning models for predicting the need for touch-up procedures following CBA in AF patients. Our findings demonstrate that a 3D deep learning model, particularly the proposed CryoAI (VoxNet++) model, can achieve promising results in preoperative assessment, potentially reducing the need for repeat procedures and improving patient outcomes. The PPVs were 96.15% for the internal validation and 95.77% for the independently external validation test sets, respectively. These results demonstrate the high diagnostic performance of the model in predicting the need for touch-up ablation. Electrophysiologists, informed by these predictions, might anticipate the potential requirement for RF touch-up ablation before the procedure, thereby facilitating alternative decision-making strategies for AF ablation. This research holds significant potential to enhance and facilitate AF CBA treatment, representing a valuable contribution to the field. Furthermore, the detailed use of 3D Activation Visualization and 3D Gaussian Curvature exemplifies explainable AI's role in identifying potential areas for touch-up, thereby improving the precision and effectiveness of treatments. 4.1. Touch-up ablation in CBA Previous research has consistently identified the left superior pulmonary vein (LSPV) and right inferior pulmonary vein (RIPV) as common sites requiring touch-up ablation in CBA procedures [14] , [26] , [27] , [28] , [29] . These findings align with the challenges noted in achieving effective PV isolation with cryoballoon alone, often due to difficulties in maintaining uniform contact. This necessitates additional RF ablations, particularly in anatomically challenging regions. Also, the association between touch-up ablation and more complex clinical profiles, including additional RF ablation outside the PV areas and higher recurrence rates, suggests a more aggressive underlying disease state [12] , [30] , [31] . The RF touch-up ablation rate in our cohort was higher than that reported in some CBA studies; however, the literature shows marked heterogeneity, with reported rates ranging from 0% to over 70%, largely influenced by verification strategies and procedural definitions [12] , [14] , [15] , [27] . For example, Wei et al. reported a 0% touch-up rate using a conservative cryoballoon-only PV isolation strategy without systematic remapping [27] , whereas studies employing rigorous post-freeze electroanatomical or voltage mapping have demonstrated substantially higher rates, including 76.6% reported by Malik et al. [12] . Other studies have shown intermediate rates of approximately 5–30% [14] , [15] , [32] . In our study, systematic mapping-guided touch-up ablation was performed whenever residual pulmonary vein conduction was detected, which likely explains the higher rate and places our findings within the broad range reported in prior literature. Investigations into balloon characteristics and anatomical factors have revealed that neither the size of the PV ostium nor the size of the cryoballoon significantly affects the need for touch-up ablation [15] . Instead, RF touch-up ablation is frequently required at anatomically thick sites, such as the anterior ridge of the LSPV, where achieving transmural lesions is complicated by the thickness of the left atrial wall relative to the cryoballoon’s energy penetration [21] . Moreover, procedural factors such as the alignment of the cryoballoon with the RIPV and anatomical constraints play significant roles in the success of the initial ablation [33] , [34] . The findings also suggest that a longer length of the PV ostium-bifurcation distance might reduce the necessity for touch-up at the RIPV, highlighting the impact of anatomical dimensions on procedural outcomes [35] . All in all, with respect to anatomical predictors, specific PV-left atrial anatomical features have been found to be associated with incomplete cryoballoon contact and a higher likelihood of touch-up ablation, including thick anterior ridges of the left superior PV [21] , steep PV-left atrial angles [33] , inferior PV orientation [34] , short PV ostium-to-bifurcation distance [35] , and increased local wall thickness [36] . These anatomically challenging regions have been consistently reported as common sites of residual conduction requiring adjunctive ablation. Our findings are concordant with and extend these prior observations. Using a 3D deep learning framework combined with Gaussian curvature analysis, we demonstrated that regions requiring touch-up ablation were consistently localized to areas of high surface curvature, which correspond to anatomically complex PV-left atrial junctions described in earlier studies. Thus, rather than relying on isolated anatomical measurements, our approach integrates global and local 3D structural information to identify high-risk regions in a data-driven manner. Beyond cryoballoon ablation, similar geometry-based modeling approaches may also have potential relevance for other ablation technologies, such as pulsed field ablation, where device-tissue conformity and anatomical configuration remain important determinants of lesion formation. Future studies may also explore whether comparable modeling frameworks could be developed using other imaging modalities, such as intracardiac echocardiography-based multiplanar imaging. The lower AUC observed in the external validation cohort compared with the internal test set likely reflects a combination of partial overfitting and genuine data distribution shifts between cohorts. CryoAI was developed and tuned using the internal cohort, in which image acquisition, reconstruction, segmentation workflows, and operator practices were relatively homogeneous. In contrast, the temporally independent validation cohort differed in case mix and procedural practice, including variability in pulmonary vein anatomical complexity, CT acquisition and segmentation parameters, and individual operators’ thresholds for performing touch-up ablation, which may attenuate model discrimination. In addition, the modest number of touch-up events makes AUC estimates sensitive to small changes in event distribution. Importantly, despite this performance drop, CryoAI retained reasonable discriminative ability on independent data, supporting its potential generalizability. ROC curves for both internal and external cohorts have been provided to allow transparent comparison of model performance across datasets in Fig. S6 . Overall, these studies collectively emphasize the critical role of anatomical factors in determining the efficacy of CBA for AF management [36] . By highlighting the specific challenges and success rates associated with different PV configurations, the research underlines the necessity of tailored procedural strategies that account for individual anatomical variations. This insight is instrumental in refining CBA techniques, ultimately enhancing patient outcomes by addressing the unique complexities presented by each case. The findings stress the importance of precise anatomical assessment in planning and executing CBA, ensuring that interventions are both effective and minimally invasive. Despite the advancements of AI in the current era of medicine [37] , there has been no application of AI in predicting the need for touch-up ablation during CBA. Our CryoAI 3D model can identify patients requiring touch-up before ablation, thereby facilitating pre-ablation decision-making. 4.2. Key findings and strengths The CryoAI (VoxNet++) model achieved superior performance compared to PointNet and VoxNet in both the internal and independent external test datasets. This suggests that VoxNet++ effectively captures global anatomical features and fine-grained local details crucial for complex 3D structures such as the anatomy of LA and PVs. The 3D Activation Visualization tool using Grad-CAM provides valuable insights into the model’s decision-making process. This interpretability allows electrophysiologists to understand the regions of interest for the AI model, potentially leading to improved trust and adoption in clinical practice. The Gaussian Curvature tool offers a complementary approach for evaluating cryoballoon placement feasibility. Identifying areas of high curvature, which may impede proper cryoballoon adherence, can enhance the prediction of locations requiring touch-up ablation. 4.3. Clinical implications This study highlights the potential of the CryoAI model for personalized preoperative assessment in CBA for AF. Early identification of complex cases requiring additional RF touch-up ablation allows for better pre-procedural planning. In this context, CryoAI is intended to function as a decision-support tool for preprocedural planning rather than a system that directly alters ablation strategy. 4.4. Limitations The study was retrospective. First, the external validation cohort was temporally independent rather than site-based, consisting of consecutive patients treated at the same institution during a later time period without overlap in patients or imaging data. Although this design allowed assessment of model robustness across different procedural eras, inter-institutional variability has not been evaluated. Second, operator-specific practice patterns, particularly regarding the threshold for performing RF touch-up ablation, may influence procedural outcomes and represent an inherent limitation of single-center studies. Accordingly, the need for RF touch-up ablation reflects procedural decision-making within the institutional workflow of our center and should not be interpreted as a purely anatomical endpoint. While the temporal validation cohort included procedures performed by multiple attending electrophysiologists, reflecting real-world variability in operator experience and decision-making, such heterogeneity was only partially addressed. Therefore, multicenter, prospective validation across diverse populations and cryoballoon types will be required to further establish the generalizability of the CryoAI model in real-time clinical settings. Third, this study did not include direct comparisons between CryoAI and clinician-based prediction or simpler feature-based machine learning models. Future studies incorporating head-to-head comparisons with expert assessment and conventional baseline models will be necessary to further clarify the added clinical value of CryoAI. Fourth, model generalizability was not assessed on large, multicenter cohorts with heterogeneous CT scanners and acquisition protocols, which may contribute to distribution shifts and partially explain the reduced performance in external validation. Future research could explore incorporating additional multimodal data sources, such as electrocardiography or fluoroscopic images, to improve model performance. 5. Conclusions This study demonstrates the promise of deep learning models, particularly CryoAI, for predicting the need for additional RF touch-up ablation in CBA for AF. Combining AI-based prediction with visualization tools and anatomical analysis using Gaussian Curvature offers a comprehensive approach to preoperative assessment. Further research and validation are warranted to translate these findings into routine clinical practice and potentially revolutionize the management of AF with CBA. CRediT authorship contribution statement Chih-Min Liu: Writing – review & editing, Writing – original draft, Visualization, Project administration, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Wei-Wen Chen: Visualization, Software, Methodology, Formal analysis. Shih-Lin Chang: Writing – review & editing, Supervision, Funding acquisition, Conceptualization. Chien-Chao Tseng: Software, Methodology. Ching-Chun Huang: Visualization, Software, Methodology, Formal analysis. Yenn-Jiang Lin: Validation, Investigation, Data curation. Li-Wei Lo: Validation, Investigation, Data curation. Yu-Feng Hu: Validation, Methodology, Investigation, Data curation. Fa-Po Chung: Validation, Methodology, Formal analysis, Data curation. Ting-Yung Chang: Validation, Investigation, Formal analysis, Data curation. Chin-Yu Lin: Validation, Investigation, Formal analysis, Data curation. Tze-Fan Chao: Validation, Investigation, Formal analysis, Data curation. Ta-Chuan Tuan: Validation, Investigation, Data curation. Jo-Nan Liao: Validation, Investigation, Data curation. Ling Kuo: Validation, Investigation, Data curation. Cheng-I Wu: Validation, Investigation, Data curation. Shin-Huei Liu: Validation, Investigation, Data curation. Jacky Chung-Hao Wu: Methodology, Formal analysis. Henry Horng-Shing Lu: Writing – review & editing, Supervision, Investigation, Formal analysis, Conceptualization. Shih-Ann Chen: Writing – review & editing, Supervision, Resources, Investigation, Conceptualization. Ethics approval and consent to participate This study was approved by the Institutional Review Board (2017–10-009BC) at Taipei Veterans General Hospital, Taipei, Taiwan. All methods were carried out following the regulations of the Institutional Review Board. The Internal Review Board of Taipei Veterans General Hospital granted an exemption from the need to secure informed consent due to the thorough de-identification of patient data. Funding This work received funding from various sources, including the National Science and Technology Council (Grants: 110-2811-M-A49-550-MY2, 110-2118-M-A49-002-MY3, 110-2314-B-075-063-MY3, 111-2634-F-A49-014-, 112-2321-B-075-002-, 112-2811-M-A49-557-, 112-2634-F-A49-003-, 113-2314-B-075-034-MY3, 113-2321-B-075A-002-, 113-2622-E-A49-010-, 113-2628-B-A49-016-, 113-2118-M-A49-007-MY2, 113-2923-M-A49-004-MY3, 114-2321-B-182-005-, 114-2622-E-A49-007-, 114-2124-M-A49-009-, 114-2124-M-037-002-), the Higher Education Sprout Project of the Kaohsiung Medical University and the National Yang Ming Chiao Tung University from the Ministry of Education, Taiwan, and the Biomedical Artificial Intelligence Academy of the Kaohsiung Medical University, Taiwan. Additional support was provided by Taipei Veterans General Hospital (Grants: VGH108C-019, VN108-12, VN109-03, V109C-070, V110C-039, V110B-043, V112B-006, V113B-008, V114C-069, V115C-103), the Ministry of Science and Technology (Grants: 108-2628-B-075-003, 109-2628-B-075-017, 109-2321-B-009-007, 109-2314-B-075-077, 110-2628-B-075-015, 110-2314-B-075-063-MY3, 110-2321-B-075-002, 110-2321-B-A49-003, 110-2634-F-A49-005-SP3, 110-2118-M-A49-002-MY3, 110-2634-F-A49-005-, and 111-2634-F-A49-014-), and Academia Sinica (Grant: AS-TM-112-01). We also thank Wan-Yi Tai for her valuable assistance and acknowledge the National Center for High-performance Computing for providing computing resources. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Footnotes Appendix A Supplementary data to this article can be found online at https://doi.org/10.1016/j.ijcha.2026.101912 . Contributor Information Shih-Lin Chang, Email: [email protected]. Henry Horng-Shing Lu, Email: [email protected]. Appendix A. Supplementary data The following are the Supplementary data to this article: Supplementary Data 1 mmc1.docx (30.4MB, docx) Data availability The data that support the findings of this study are available from the corresponding author on request. References 1. Linz D., Gawalko M., Betz K., Hendriks J.M., Lip G.Y.H., Vinter N., et al. Atrial fibrillation: epidemiology, screening and digital health. Lancet Reg Health Eur. 2024;37 doi: 10.1016/j.lanepe.2023.100786. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 2. Chao T.F., Liu C.J., Tuan T.C., Chen T.J., Hsieh M.H., Lip G.Y.H., et al. Lifetime risks, projected numbers, and adverse outcomes in asian patients with atrial fibrillation: a report from the Taiwan Nationwide AF Cohort Study. Chest. 2018;153:453–466. doi: 10.1016/j.chest.2017.10.001. [ DOI ] [ PubMed ] [ Google Scholar ] 3. Tzeis S., Gerstenfeld E.P., Kalman J., Saad E.B., Sepehri Shamloo A., Andrade J.G., et al. European heart rhythm association/heart rhythm society/asia pacific heart rhythm society/latin american heart rhythm society expert consensus statement on catheter and surgical ablation of atrial fibrillation. Europace. 2024;2024:26. [ Google Scholar ] 4. Joglar J.A., Chung M.K., Armbruster A.L., Benjamin E.J., Chyou J.Y., Cronin E.M., et al. 2023 ACC/AHA/ACCP/HRS guideline for the diagnosis and management of atrial fibrillation: a report of the american college of cardiology/american heart association joint committee on clinical practice guidelines. Circulation. 2024;149:e1–e156. doi: 10.1161/CIR.0000000000001193. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Chen S., Pürerfellner H., Ouyang F., Kiuchi M.G., Meyer C., Martinek M., et al. Catheter ablation vs. antiarrhythmic drugs as 'first-line' initial therapy for atrial fibrillation: a pooled analysis of randomized data. Europace. 2021;23:1950–1960. doi: 10.1093/europace/euab185. [ DOI ] [ PubMed ] [ Google Scholar ] 6. Wazni O.M., Dandamudi G., Sood N., Hoyt R., Tyler J., Durrani S., et al. Cryoballoon ablation as initial therapy for atrial fibrillation. N. Engl. J. Med. 2021;384:316–324. doi: 10.1056/NEJMoa2029554. [ DOI ] [ PubMed ] [ Google Scholar ] 7. Maurer T., Schlüter M., Kuck K.H. Keeping it simple: balloon devices for atrial fibrillation ablation therapy. JACC Clin. Electrophysiol. 2020;6:1577–1596. doi: 10.1016/j.jacep.2020.08.041. [ DOI ] [ PubMed ] [ Google Scholar ] 8. Garan A., Al-Ahmad A., Mihalik T., Cartier C., Capuano L., Holtan D., et al. Cryoablation of the pulmonary veins using a novel balloon catheter. J. Interv. Card. Electrophysiol. 2006;15:79–81. doi: 10.1007/s10840-006-8519-5. [ DOI ] [ PubMed ] [ Google Scholar ] 9. Hoffmann E., Straube F., Wegscheider K., Kuniss M., Andresen D., Wu L.Q., et al. Outcomes of cryoballoon or radiofrequency ablation in symptomatic paroxysmal or persistent atrial fibrillation. Europace. 2019;21:1313–1324. doi: 10.1093/europace/euz155. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Mörtsell D., Arbelo E., Dagres N., Brugada J., Laroche C., Trines S.A., et al. Cryoballoon vs. radiofrequency ablation for atrial fibrillation: a study of outcome and safety based on the ESC-EHRA atrial fibrillation ablation long-term registry and the swedish catheter ablation registry. Europace. 2019;21:581–589. doi: 10.1093/europace/euy239. [ DOI ] [ PubMed ] [ Google Scholar ] 11. Andrade J.G., Khairy P., Dubuc M. Catheter cryoablation: biology and clinical uses. Circ. Arrhythm. Electrophysiol. 2013;6:218–227. doi: 10.1161/CIRCEP.112.973651. [ DOI ] [ PubMed ] [ Google Scholar ] 12. Malik R., Malik B., Hunter T.D. Touch-up and recurrence rates after voltage mapping for verification of pulmonary vein isolation following cryoablation of paroxysmal atrial fibrillation. J. Interv. Card. Electrophysiol. 2019;56:307–312. doi: 10.1007/s10840-019-00533-y. [ DOI ] [ PubMed ] [ Google Scholar ] 13. Matsunaga Y., Egami Y., Yano M., Yamato M., Shutta R., Nishino M., et al. Clinical outcome of non-use of touch-up focal ablation catheters strategy during cryoballoon atrial fibrillation ablation. Eur. Heart J. 2020;41 [ Google Scholar ] 14. Miyazaki S., Taniguchi H., Hachiya H., Nakamura H., Takagi T., Hirao K., et al. Clinical recurrence and electrical pulmonary vein reconnections after second-generation cryoballoon ablation. Heart Rhythm. 2016;13:1852–1857. doi: 10.1016/j.hrthm.2016.05.025. [ DOI ] [ PubMed ] [ Google Scholar ] 15. Wakamatsu Y., Nagashima K., Nakahara S., Iso K., Watanabe R., Arai M., et al. Electrophysiologic and anatomic factors predictive of a need for touch-up radiofrequency application for complete pulmonary vein isolation: Comparison between hot balloon- and cryoballoon-based ablation. J. Cardiovasc. Electrophysiol. 2019;30:1261–1269. doi: 10.1111/jce.13989. [ DOI ] [ PubMed ] [ Google Scholar ] 16. Sánchez-Cruz H., Sossa-Azuela H., Braumann U.D., Bribiesca E. The Euler-Poincaré formula through contact surfaces of voxelized objects. J. Appl. Res. Technol. 2013;11:65–78. [ Google Scholar ] 17. Aryana A., Kenigsberg D.N., Kowalski M., Koo C.H., Lim H.W., O'Neill P.G., et al. Verification of a novel atrial fibrillation cryoablation dosing algorithm guided by time-to-pulmonary vein isolation: results from the Cryo-DOSING Study (Cryoballoon-ablation DOSING based on the Assessment of Time-to-effect and Pulmonary Vein Isolation Guidance) Heart Rhythm. 2017;14:1319–1325. doi: 10.1016/j.hrthm.2017.06.020. [ DOI ] [ PubMed ] [ Google Scholar ] 18. Liu C.M., Chen W.S., Chang S.L., Hsieh Y.C., Hsu Y.H., Chang H.X., et al. Use of artificial intelligence and I-Score for prediction of recurrence before catheter ablation of atrial fibrillation. Int. J. Cardiol. 2024;402 doi: 10.1016/j.ijcard.2024.131851. [ DOI ] [ PubMed ] [ Google Scholar ] 19. Liu C.M., Chang S.L., Chen H.H., Chen W.S., Lin Y.J., Lo L.W., et al. The clinical application of the deep learning technique for predicting trigger origins in patients with paroxysmal atrial fibrillation with catheter ablation. Circ. Arrhythm. Electrophysiol. 2020;13 doi: 10.1161/CIRCEP.120.008518. [ DOI ] [ PubMed ] [ Google Scholar ] 20. Meyer M., Desbrun M., Schröder P., Barr A.H. In: Visualization and Mathematics III. Hege H.-.-C., Polthier K., editors. Berlin, Heidelberg; Springer, Berlin Heidelberg: 2003. Discrete Differential-Geometry Operators for Triangulated 2-Manifolds; pp. 35–57. [ Google Scholar ] 21. Iso K., Okumura Y., Watanabe I., Nagashima K., Sonoda K., Kogawa R., et al. Wall thickness of the pulmonary vein-left atrial junction rather than electrical information as the major determinant of dormant conduction after contact force-guided pulmonary vein isolation. J. Interv. Card. Electrophysiol. 2016;46:325–333. doi: 10.1007/s10840-016-0147-0. [ DOI ] [ PubMed ] [ Google Scholar ] 22. R.Q. Charles, H. Su, M. Kaichun, L.J. Guibas, PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)2017. p. 77–85. 23. D. Maturana, S. Scherer, VoxNet: A 3D Convolutional Neural Network for real-time object recognition. 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)2015. p. 922–928. 24. N. Sedaghat, M. Zolfaghari, T. Brox, Orientation-boosted Voxel Nets for 3D Object Recognition. ArXiv. 2016;abs/1604.03351. 25. R.R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, D. Batra, Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. 2017 IEEE International Conference on Computer Vision (ICCV)2017. p. 618–626. 26. Kühne M., Suter Y., Altmann D., Ammann P., Schaer B., Osswald S., et al. Cryoballoon versus radiofrequency catheter ablation of paroxysmal atrial fibrillation: biomarkers of myocardial injury, recurrence rates, and pulmonary vein reconnection patterns. Heart Rhythm. 2010;7:1770–1776. doi: 10.1016/j.hrthm.2010.08.028. [ DOI ] [ PubMed ] [ Google Scholar ] 27. Wei H.Q., Guo X.G., Zhou G.B., Sun Q., Liu X., Yang J.D., et al. Pulmonary vein isolation with real-time pulmonary vein potential recording using second-generation cryoballoon: Procedural and biophysical predictors of acute pulmonary vein reconnection. Pac. Clin. Electrophysiol. 2018;41:14–21. doi: 10.1111/pace.13230. [ DOI ] [ PubMed ] [ Google Scholar ] 28. Ciconte G., Chierchia G.B., De Asmundis C., Sieira J., Conte G., Juliá J., et al. Spontaneous and adenosine-induced pulmonary vein reconnection after cryoballoon ablation with the second-generation device. J. Cardiovasc. Electrophysiol. 2014;25:845–851. doi: 10.1111/jce.12421. [ DOI ] [ PubMed ] [ Google Scholar ] 29. Miki K., Fukuda K., Hirano M., Sato K., Ikeda S., Shinozaki M., et al. Localization and spread of challenging conduction gaps of pulmonary veins for atrial fibrillation cryoablation. Pac. Clin. Electrophysiol. 2025;48:21–29. doi: 10.1111/pace.15133. [ DOI ] [ PubMed ] [ Google Scholar ] 30. Wójcik M., Berkowitsch A., Zaltsberg S., Hamm C.W., Pitschner H.F., Kuniss M., et al. Cryoballoon ablation of atrial fibrillation: how important is the proper selection of patients? Cardiol. J. 2015;22:194–200. doi: 10.5603/CJ.a2014.0100. [ DOI ] [ PubMed ] [ Google Scholar ] 31. Akkaya E., Berkowitsch A., Greiss H., Hamm C.W., Sperzel J., Neumann T., et al. PLAAF score as a novel predictor of long-term outcome after second-generation cryoballoon pulmonary vein isolation. Europace. 2018;20:f436–f443. doi: 10.1093/europace/eux295. [ DOI ] [ PubMed ] [ Google Scholar ] 32. Kimura M., Kobori A., Nitta J., Hirao K., Shizuta S., Kurita T., et al. Cryoballoon ablation for paroxysmal atrial fibrillation in Japan: 2-year safety and efficacy results from the Cryo AF Global Registry. J. Interv. Card. Electrophysiol. 2022;64:695–703. doi: 10.1007/s10840-022-01132-0. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 33. Yasuoka R., Kurita T., Kotake Y., Hashiguchi N., Motoki K., Kobuke K., et al. Particular morphology of inferior pulmonary veins and difficulty of cryoballoon ablation in patients with paroxysmal atrial fibrillation. Circ J. 2017;81:668–674. doi: 10.1253/circj.CJ-16-1161. [ DOI ] [ PubMed ] [ Google Scholar ] 34. Knecht S., Kühne M., Altmann D., Ammann P., Schaer B., Osswald S., et al. Anatomical predictors for acute and mid-term success of cryoballoon ablation of atrial fibrillation using the 28 mm balloon. J. Cardiovasc. Electrophysiol. 2013;24:132–138. doi: 10.1111/jce.12003. [ DOI ] [ PubMed ] [ Google Scholar ] 35. Kajiyama T., Miyazaki S., Matsuda J., Watanabe T., Niida T., Takagi T., et al. Anatomic parameters predicting procedural difficulty and balloon temperature predicting successful applications in individual pulmonary veins during 28-mm second-generation cryoballoon ablation. JACC Clin. Electrophysiol. 2017;3:580–588. doi: 10.1016/j.jacep.2017.01.004. [ DOI ] [ PubMed ] [ Google Scholar ] 36. Hayashi T., Murakami M., Saito S., Iwasaki K. Characteristics of anatomical difficulty for cryoballoon ablation: insights from CT. Open. Heart. 2022;9 doi: 10.1136/openhrt-2021-001724. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 37. Kabra R., Israni S., Vijay B., Baru C., Mendu R., Fellman M., et al. Emerging role of artificial intelligence in cardiac electrophysiology. Cardiovasc Digit Health J. 2022;3:263–275. doi: 10.1016/j.cvdhj.2022.09.001. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary Data 1 mmc1.docx (30.4MB, docx) Data Availability Statement The data that support the findings of this study are available from the corresponding author on request. Articles from International Journal of Cardiology. Heart & Vasculature are provided here courtesy of Elsevier ACTIONS View on publisher site PDF (2.7 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top