Predicting Cerebral Aneurysm Recurrence after Coil Embolization: A Novel Deep Learning Approach Using Time-of-flight...

The recurrence of cerebral aneurysms after coil embolization remains a significant concern in clinical practice. This study introduced a novel approach that combines machine learning with deep learning techniques using time-of-flight magnetic resonance angiography to predict aneurysm recurrence. A retrospective multicenter analysis was conducted on 154 patients with coil-embolized unruptured cerebral aneurysms. Three prediction models were developed: a logistic regression model, a neural network model using clinical data, and a combined deep learning model incorporating both clinical and imaging data obtained from 3-dimensional reconstructed time-of-flight magnetic resonance angiography. The combined model was created in 2 versions: 1 trained exclusively with pre-operative images and the other using both pre- and post-operative images. All models were evaluated using leave-one-out cross-validation to assess the area under the receiver operating characteristic curve, accuracy, sensitivity, and specificity. Among the 154 cases, 46 (29.9%) demonstrated recurrence. The combined model that incorporated both pre- and post-operative time-of-flight magnetic resonance angiography images achieved the best discriminative performance, with an area under the receiver operating characteristic curve of 0.822 and sensitivity of 0.848, indicating a reliable capacity to identify patients at risk of recurrence. The version that used only pre-operative images showed relatively high accuracy (0.766) and specificity (0.833) but a lower overall area under the receiver operating characteristic curve. These results highlight the value of integrating time-of-flight magnetic resonance angiography data into artificial intelligence-driven predictive models to improve the assessment of recurrence risk after coil embolization. The improved performance of the combined model, particularly when including post-operative images, demonstrates potential for more precise, individualized risk prediction and may contribute to better patient management.