A generative adversarial framework for optimal view prediction in aneurysm embolization

Selecting an optimal C-arm working view is critical in endovascular coiling of intracranial aneurysms, influencing procedural safety and efficiency. Current practice relies on operator experience and manual angulation adjustments, leading to inter-operator variability and increased radiation exposure. In this preliminary feasibility study, we investigate whether a generative adversarial framework can learn to predict clinically meaningful C-arm viewing directions from 3D rotational angiography (3DRA) data.

We developed a generator-discriminator framework in which the generator predicts a 3D unit view vector from segmented vascular and aneurysm volumes, and the discriminator evaluates differentiable 2D projections along the predicted view. Three projection strategies (soft first-hit ray casting, digitally reconstructed radiographs (DRR), and label-aware maximum-intensity projections with soft label encoding (MIP-OR)) were tested with two generator architectures (CNN and U-Net). Performance was assessed on a held-out test set using absolute dot product (ADP) error between predicted and expert-annotated views, complemented by qualitative assessment from two experienced interventional neuroradiologists.

Experiments on a preliminary cohort of 18 patients demonstrate that the CNN generator combined with MIP-OR projections achieves the lowest mean ADP error and high expert scores, demonstrating improved alignment with clinically acceptable working views. Training analysis indicated that low ADP values alone do not guarantee clinical usability, highlighting the need for combined geometric and expert-based evaluation.

Adversarial learning with anatomically informed projections is a promising approach for automated C-arm view prediction. Future work will integrate procedural ground-truth views, larger datasets, and additional endovascular procedures to evaluate reliability and generalizability.