Abstract 150: Development and validation of a deep learning algorithm to predict coil morphology change and aneurysm recurrence on skull X‐ray after coiling using a retrospective cohort

Abstract

Introduction Digital subtraction angiography (DSA) is the gold standard method to screen for aneurysm recurrence after coil embolization, but it is invasive and resource intensive. Computed tomographic angiography and magnetic resonance arteriography can screen for aneurysm recurrence but these modalities are limited by coil artifact. Deep learning algorithm analysis of skull X‐rays could potentially predict aneurysm recurrence and help select patients requiring evaluation with DSA. Methods We conducted a retrospective, single center cohort study using a database consisting of aneurysms treated with coiling between 2008 and 2024 (n=170). We obtained initial postoperative skull X‐rays and follow up skull X‐rays at 6, 12, 24, and/or 60 months. We trained a U‐Net based coil segmentation model to localize the coil mass and its masks using 746 manually annotated skull X‐rays. Skull X‐ray image pairs (postoperative and follow‐up images) were combined into two‐channel inputs for a DenseNet‐264 based model to train a coil change classification model using 117 patients with 363 x‐ray pairs. Ground truth was determined by core lab physician rating of skull X‐rays. The coil change classification model outputs the probability of coil morphology change between timepoints for the anterior‐posterior (AP) and lateral (LT) X‐rays. Composite case‐level predictions were defined as the maximum probability across AP and LT views. We selected a prediction threshold of 0.48 to optimize for high recall. We validated the coil change classification model with a cohort of 53 patients with 145 X‐ray pairs. Ground truth was determined by core lab physician rating of DSAs. Results The validation cohort consisted of 69 cases, 53 patients, and 145 X‐ray pairs. 17 patients (24.6%) experienced aneurysm recurrence confirmed by DSA. The sensitivity was 94.1%, specificity was 32.7%, accuracy was 47.8%, precision was 31.4%, and recall was 94.1%. Conclusion The described deep learning algorithm performed well in predicting coil morphology change in coiled aneurysms on skull X‐rays. With additional training, this algorithm could be developed into a cost‐effective method to screen for aneurysm recurrence and guide selection of patients who require evaluation with DSA.

Ähnliche Arbeiten