Prediction of cerebral aneurysm rupture using a point cloud neural network

Area of Science:

• Neurosurgery
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate prediction of cerebral aneurysm (CA) rupture is critical for patient management.
• Current prediction models often rely on radiomics features extracted from imaging data.

Purpose of the Study:

• To evaluate the accuracy of a point cloud neural network (PC-NN) in predicting CA rupture.
• To compare PC-NN performance against traditional radiomics-based models using MR angiography (MRA) and CT angiography (CTA) data.

Main Methods:

• Retrospective analysis of 418 CAs from 411 patients, using CTA or MRA data.
• Development and comparison of a PC-NN model with ridge regression, SVM, and NN radiomics models.
• Prospective evaluation of PC-NN and radiomics models in 258 CAs from five external centers.

Main Results:

• The PC-NN model demonstrated significantly higher accuracy (AUC=0.913) than radiomics models (AUCs 0.788-0.805) in internal testing.
• PC-NN models trained with MRA data (AUC=0.936) showed superior performance compared to CTA data (AUC=0.824).
• In external validation, PC-NN achieved a higher AUC (0.835) than radiomics models (AUCs 0.681-0.701).

Conclusions:

• PC-NNs offer a more accurate approach to predicting cerebral aneurysm rupture than conventional radiomics models.
• MRA-based PC-NN models exhibit superior predictive performance compared to CTA-based models.