Use of a convolutional neural network for aneurysm identification in digital subtraction angiography

Angiographic Parametric Imaging (API) is a quantitative image analysis method that uses a digital subtraction angiography (DSA) to characterize contrast media dynamics throughout vasculature. The parameters acquired through API may be used to assess the success of a neurovascular intervention such as the stenting or coiling of an aneurysm. This imaging tool requires manual contouring of the aneurysm sac and the surrounding vasculature, which is not realistic in an interventional suite. To address this challenge, we studied whether convolutional neural networks can carry out a three-class segmentation problem differentiating between the background, vasculature, and aneurysm sac in a DSA acquisition. Image data were retrospectively collected from patients being monitored or treated for cerebral aneurysms at Gates Vascular Institute. While VGG-16 and U-NET architecture were both investigated, a modified VGG architecture was developed and used. Network training was carried out over 100 epochs. Our training dataset comprised of 12000 DSA acquisitions. Our validation dataset comprised of 2000 DSA acquisitions. The Jaccard Index was above 0.74 for both classes. The Dice similarity coefficient was above 0.83 for both classes. Area under the ROC curve was above 0.72 for both classes. These results indicate good agreement between the ground-truth labels and the network predicted labels. Our network proved not sensitive to motion artifacts or the presence of skull in the image data. This work indicates the potential clinical utility of a convolutional neural network in the context of aneurysm detection in DSA for feature extraction using parametric imaging to support a clinical decision.