Learning-based automatic segmentation on arteriovenous malformations from contract-enhanced CT images

We propose a learning-based method to automatically segment arteriovenous malformations (AVM) target volume from computed tomography (CT) in stereotactic radiosurgery (SRS). A deeply supervised 3D V-Net is introduced to enable end-to-end segmentation. Deep supervision mechanism is integrated into the hidden layers to overcome the optimization difficulties when training such a network with limited training data. The probability map of new AVM contour is generated by the well-trained network. To evaluate the proposed method, we retrospectively investigate 30 AVM patients treated by SRS. For each patient, both digital subtraction angiography (DSA) and CT with contrast had been acquired. Using our proposed method, the AVM contours are generated solely based on contrast CT images, and are compared with the AVM contours delineated from DSA by physicians as ground truth. The average centroid distance, volume difference and DSC value among all 30 patients are 0.83±0.91mm, -0.01±0.79 and 0.84±0.09, which indicates that the proposed method is able to generate AVM target contour with around 1mm error in displacement, 1cc error in volume size and 84% overlapping compared with ground truth. The proposed method has great potential in eliminating DSA acquisition and developing a solely CT-based treatment planning workflow for AVM SRS treatment.