Abstract WP269: Automatic Segmentation of Cerebral Arteries in MRA TOF Using Deep Learning

Purpose: Time-of-Flight (TOF) MRA is commonly used for grading cerebrovascular diseases. Analysis of cerebral arteries in MRA TOF is a challenging and time consuming task that would benefit from automation. Established image processing methods for automatic segmentation of cerebral arteries suffer from common artefacts such as kissing vessels (when two nearby vessels touch) and signal intensity drop (especially in the presence of pathology). Artificial intelligence models are promising candidates for resolving such artefacts. Here, we propose and assess the performance of a deep learning model for automatic segmentation of cerebral arteries in MRA TOF which is robust to common MRI artefacts. Methods: A 3D convolutional neural network (CNN) is proposed for automatic segmentation of intracranial arteries in MRA TOF. The neural network is trained with a custom loss function and residual blocks to penalize the occurrence of common artefacts such as kissing vessels. The model is trained and tested on a dataset consisting of 82 subjects (50 healthy volunteers and 32 patients with intracranial stenosis) following a 3-fold cross-validation method, i.e. 3 models are trained where each model is blind to one-third of the data in the training process to avoid bias. Manual segmentation of the arteries done by an expert reader are used as ground-truth for training and testing the model. Results: The proposed deep learning model achieved excellent accuracy compared against the ground truth (Dice score 0.89). Our proposed deep learning model outperformed a state-of-the-art neural network for image segmentation (3DU-Net, Dice score 0.85) and resulted in considerably less occurences of artefacts such as kissing vessels (9% of cases had segmentation artefacts for our model vs 16% for 3D U-Net). The proposed deep learning model was fast, taking only 2 seconds to produce a 3D model of the arteries on a laptop with a dedicated GPU. Conclusion: The proposed deep learning model accurately segments intracranial arteries in MRA TOF and is robust to common artefacts of MR imaging thanks to implementation of a custom loss function. The model can potentially increase the accuracy and speed of grading cerebrovascular diseases.