Deep learning-based fine-grained assessment of aneurysm wall characteristics using 4D-CT angiography

This study proposes a novel deep learning-based approach for aneurysm wall characteristics, including thin-walled (TW) and hyperplastic-remodeling (HR) regions.

We analyzed fifty-two unruptured cerebral aneurysms employing 4D-computed tomography angiography (4D-CTA) and intraoperative recordings. The TW and HR regions were identified in intraoperative images. The 3D trajectories of observation points on aneurysm walls were processed to compute a time series of 3D speed, acceleration, and smoothness of motion, aiming to evaluate the aneurysm wall characteristics. To facilitate point-level risk evaluation using the time-series data, we developed a convolutional neural network (CNN)-long- short-term memory (LSTM)-based regression model enriched with attention layers. In order to accommodate patient heterogeneity, a patient-independent feature extraction mechanism was introduced. Furthermore, unlabeled data were incorporated to enhance the data-intensive deep model.

The proposed method achieved an average diagnostic accuracy of 92%, significantly outperforming a simpler model lacking attention. These results underscore the significance of patient-independent feature extraction and the use of unlabeled data.

This study demonstrates the efficacy of a fine-grained deep learning approach in predicting aneurysm wall characteristics using 4D-CTA. Notably, incorporating an attention-based network structure proved to be particularly effective, contributing to enhanced performance.