Abstract Background Aortic dissection (AD) is a life-threatening disease. Artificial intelligence (AI) offers promising solutions for automating morphological analysis and improving predictive accuracy. However, conventional ML models often underutilize the spatial complexity of aortic geometry. This study addresses these gaps by leveraging point cloud neural networks (PC-NNs), which excel in capturing intricate 3D structural features, to predict TBAD occurrence. Purpose This study aimed to develop a PC-NN framework to automatically analyze 3D aortic morphology from computed tomography angiography (CTA) data and compare the performance of PC-NN against conventional ML methods in predicting TBAD. Methods Study Design and Participants. A retrospective multicenter study enrolled 413 participants from a provincial hospital and a university hospital (January 2018–August 2024). The cohort included 161 TBAD patients and 252 non-TBAD controls. Inclusion criteria for TBAD patients required confirmed Stanford B dissection via CTA, age 18–80 years, and no prior aortic/cardiac surgery. Non-TBAD patients had similar inclusion criteria but lacked aortic pathology. Model Development 1. Traditional ML Models: Extracted geometric features included aortic diameters, lengths and angulation metrics. Four ML algorithms (Random Forest, XGBoost, LightGBM, CatBoost) were trained on 70% of the data (30% testing). 2. PC-NN Architecture: Based on PointNet++, Training utilized the Adam optimizer on an NVIDIA RTX 4060 GPU. Results Traditional ML Performance XGBoost outperformed other ML models, achieving an AUC of 0.877, accuracy of 0.784, sensitivity of 0.833, and specificity of 0.854. However, sensitivity remained suboptimal for clinical deployment. PC-NN Superiority AUC-ROC: 0.950 (95% CI: 0.934–0.986) Sensitivity: 0.868 Specificity: 0.967 Accuracy: 0.881 Key Advantages of PC-NN 1. Automated Feature Extraction: Eliminated manual measurement biases and captured subtle spatial patterns (e.g., localized tortuosity, asymmetric dilation). 2. Multi-Scale Learning: Integrated local (e.g., aortic arch curvature) and global (e.g., ascending-descending length ratio) features. 3. Generalizability: Achieved high reproducibility across heterogeneous CTA datasets. Conclusion This study establishes PC-NN as a transformative tool for TBAD prediction, surpassing conventional ML methods in accuracy, sensitivity, and specificity. By automating 3D aortic morphological analysis, PC-NN mitigates subjective measurement errors and uncovers latent spatial risk patterns. The findings underscore the clinical potential of AI-driven risk stratification, particularly for high-risk populations undergoing routine CTA surveillance. Future work should focus on prospective validation and integration into real-time diagnostic workflows to enhance early detection and preventive care for aortic dissection.Point cloud network structure Performance of ML and PC-NN
Cheng et al. (Sat,) studied this question.