Three-dimensional deep neural network integrating transfer learning for preoperative coronary CTA classification in atrial fibrillation patients(PDF)
《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]
- Issue:
- 2025年第9期
- Page:
- 1245-1254
- Research Field:
- 医学人工智能
- Publishing date:
Info
- Title:
- Three-dimensional deep neural network integrating transfer learning for preoperative coronary CTA classification in atrial fibrillation patients
- Author(s):
- CHEN Wei1; 2; 3; 4; XIN Zirui1; 3; 4; CHEN Xi1; 4; LIU Zhenjiang1; 4; LUO Aijing1; 2; 3; 4
- 1. The Second Xiangya Hospital of Central South University, Changsha 410011, China 2. School of Life Sciences, Central South University, Changsha 410013, China 3. Key Laboratory of Medical Information Research in Hunan Province (Central South University), Changsha 410013, China 4. Clinical Medical Research Center for Cardiovascular Intelligent Medicine in Hunan Province, Changsha 410011, China
- Keywords:
- Keywords: atrial fibrillation coronary computed tomography angiography three-dimensional deep neural network transfer learning catheter ablation surgical strategy
- PACS:
- R318；R541.75
- DOI:
- DOI:10.3969/j.issn.1005-202X.2025.09.017
- Abstract:
- Abstract: Objective To develop a three-dimensional (3D) deep neural network based preoperative classification model for coronary computed tomography angiography (CTA) in atrial fibrillation patients, and to explore the effects of transfer learning on the performance of medical image classification models, thereby providing preoperative decision support for catheter ablation to advance atrial fibrillation treatment toward precision and personalization. Methods Utilizing 3D ConvNet and 3D ResNet as backbone network, the three-dimensional classification features were extracted from coronary CTA sequences. The publicly available pre-trained weights were used for transfer learning. The model performance was evaluated through metrics such as confusion matrix, classification accuracy, and area under the curve (AUC). A comparative analysis was also conducted to evaluate the performance differences between the transfer learning model and the initialized training model. Results Transfer learning yielded significant performance improvements over the initialized training models, attaining AUC improvement of 9.1%-16.7% and accuracy enhancement of 6.2%-23.5%. Among all models, 3D-ResNet18 model with MedicalNet pre-training weights performed the best, achieving an AUC of 0.77 and an accuracy of 0.71. Conclusion The proposed three-dimensional deep network enhanced by transfer learning can effectively identify atrial fibrillation patients requiring additional ablation besides pulmonary vein isolation through preoperative coronary CTA, which will assist clinicians in optimizing surgical strategies and improving treatment outcomes, thereby reducing long-term postoperative recurrence rates.
Last Update: 2025-09-30