Automatic tools to identify aneurysms and predict long-term treatment outcomes
Mohammad Bhurwani, Kenneth Snyder, Alexander Podgorsak, Ciprian Ionita, Adnan Siddiqui, Jason Davies, Elad Levy
Introduction: Digital subtraction angiography (DSA) is an essential tool in the diagnosis and treatment of aneurysms, but its use is largely qualitative. Angiographic parametric imaging (API) is a tool that quantifies hemodynamic factors, known to be important in the initiation, growth, rupture, and healing of aneurysms.
Objective: We investigated the feasibility of creating a computational pipeline using machine learning to automate aneurysm detection and treatment outcome prediction.
Methods: Three hundred fifty angiograms of pre- and post-coiled IAs were used to train a convolutional neural network to automatically segment aneurysms using manually contoured aneurysm masks. The output was used to automatically compute API within the aneurysm. Using an additional fourty-eight patients with fifty-two treated aneurysms, we collected angiograms pre-treatment, post-treatment and at six months and trained a deep neural network to predict the long-term treatment outcome using intraoperative angiography.
Results: The algorithm was able to accurately identify aneurysms and compute API comparable to humans. The mean Jaccard Index was 0.831 for aneurysms and 0.744 for parent vessels, with mean Dice similarity coefficient of 0.905. The algorithm predicting treatment outcomes had AUROC of 0.866, indicating excellent performance. Evaluating performance based on whether treatment was undertaken with coiling, flow-diversion, and combined treatment yielded ROC of 0.79, 0.86, and 0.64, respectively.
Conclusions: Using deep learning methods and API, we can accurately identify aneurysms and predict treatment outcomes at the time of treatment across a range of treatment types. Such algorithms could help to inform therapeutic decisions and improve outcomes for aneurysm treatment.