3D Segmentation of Abdominal Aortic Aneurysm Walls from CT Angiograms

Abdominal aortic aneurysm (AAA) wall segmentation from computed tomography angiography (CTA) imaging provides an understanding of the disease’s size, location, and geometry for endovascular repair. Segmentation can also serve as an intermediate to wall stress computation associated with biomechanical evaluation. However, the task of segmenting thin, annular objects can be challenging due to the high resolution and large memory requirements for training convolutional neural networks (CNNs). We engineer a CNN based on the U-Net architecture for patch-wise 3D segmentation. The model infers upon CTA sub-volumes before reconstituting the predictions into the full AAA wall segmentation. The study cohort comprised of 123 patient CTAs presenting for symptomatic and asymptomatic cases. The training set consisted of 73 patients (38 Asymptomatic & 35 Symptomatic) while the test set consisted of 50 patients (25 Asymptomatic & 25 Symptomatic). After preparing the model inputs of size 256X256X64 voxels, the model was trained using sub-volumes of size 128X128X64 voxels and fit according to the Generalized Dice Loss. We determine accuracy via the dice score coefficient (DSC), comparing predicted segmentations to ground truths acquired at the original frame of reference. For the asymptomatic test set, the model achieves DSCs of mean 0.42 and median 0.51. For the symptomatic test set, the model achieves DSCs of mean 0.32 and median 0.33. The model could accelerate time spent in preparatory steps prior to biomechanical modeling of wall stress or provide clinicians with rapid evaluation of AAA morphology to stratify rupture risk.