Abstract 15889: Evaluation of a Novel CTA/DSA-Based Deep-Learning Cerebrovascular Segmentation Method for Cavernous Intracranial Aneurysms

Abstract

Introduction: CTA is a mainstay diagnostic imaging modality for screening and follow-up of intracranial aneurysms (IAs). One challenge facing neuroradiologists is detection and visualization of IAs on the cavernous internal carotid artery (ICA), which are often obstructed by bony tissue. Such IAs are either missed diagnosed, or must be re-imaged by high-resolution DSA. We sought to evaluate the performance of automated deep learning methods in segmenting cavernous ICA aneurysms from CTA. Methods: In 30 cases, we collected and registered CTA-DSA image pairs to train 2 deep-learning (DL) networks (namely, Brave-Net [BN], and DeepMedic [DM]) for brain vessel segmentation from CTA images of patients with IAs. Segmented DSA was used as the ground truth. Independent testing was performed on a separate cohort of 20 cases to compare network performances. The best-performing network was used for comparison against expert rater assessment in IA detection and segmentation. Quantitative evaluation of expert rater IA segmentation was also performed by comparing IA morphology quantified by parameters, such as size, aspect ratio, neck diameter, and undulation index, against DL predictions. Results: In the independent testing dataset, BN significantly outperformed DM in terms of Dice similarity coefficient DSC (0.875±0.0083 vs. 0.864±0.0075, p=0.004), while DM significantly outperformed BN in terms of precision (0.891±0.016 vs. 0.867±0.017, p<0.001). Qualitative assessment confirmed the superiority of DM in segmenting the IAs located on cavernous ICA. Compared to DM, expert raters missed 7 IAs located on cavernous ICA. Furthermore, DM segmentation provided better measurement of neck diameter (p=0.009), aspect ratio (p=0.029) and shape parameters as compared to the expert rater assessment. Conclusions: A deep learning approach using DM can provide a high-fidelity solution for CTA segmentation of vessels and IAs in the cavernous ICA region.