One shot lumen mesh generation of abdominal aortic aneurysm by hybrid neural network

Abstract

BACKGROUND: The majority of current algorithms for blood flow surface extraction in the context of hemomodeling of abdominal aortic aneurysms are derived through a segmentation step, rather than directly from CT scans [1]. This approach introduces a degree of complexity, as the segmentation neural network is trained without consideration of the fact that the blood flow is a simply-connected region. Consequently, post-processing may be required to fulfill the simple connectivity criterion. In addition, the blood flow surface obtained from the segmentation mask using marching cubes is too coarse and requires smoothing. To provide one-stage surface extraction, Voxel2Mesh [2] was the first to be proposed. Voxel2Mesh shows good performance in extracting relatively simple geometries, while for more complex ones, its modifications have been proposed in the literature [3, 4]. AIM: The study aimed to develop an algorithm for single-stage extraction of the lumen surface of an abdominal aortic aneurysm. MATERIALS AND METHODS: A total of 90 contrast-enhanced CT images and segmentation masks with blood flow region labeling were prepared and divided into three groups: 40, 20, and 30 images for training, validation, and testing, respectively. Affine and non-linear augmentations were applied to increase the effective training sample size. A hybrid neural network consisting of a voxel encoder, a voxel decoder, and a grid decoder was proposed for single-stage surface extraction. The architectural design of the encoder is based on the Atto-sized ConvNeXtV2 architecture. The voxel decoder is comprised of five blocks, beginning with an interpolation layer and concluding with two super-precision words with packet normalization layers and ReLU. The voxel decoder and encoder are linked by means of analogous connections to those observed in the Unet architecture. The grid decoder comprises four GraphSAGE convolutions, with GeLU intervening between each pair. It is connected to the voxel decoder. The input to the encoder is a computed tomography image, while the input to the grid decoder is an initial approximation of the surface in the form of a ball. The output of the voxel decorrelation is a segmentation mask, while the output of the mesh decorrelation is the extracted surface. A combination of voxel and mesh loss functions was employed for the purposes of training. The surface generated from the segmentation mask by the Marching Cubes algorithm was employed as the reference surface. The mesh loss function was regularized to set the necessary parameters for the generated mesh. The quality of the generated mesh was evaluated using the Dice coefficient, which compares the true segmentation mask with the rasterized generated surface. RESULTS: We proposed the first hybrid neural network with an encoder based on the state-of-the-art ConvNeXtV2 architecture for the direct generation of abdominal aortic aneurysm blood flow meshes. A 14.01% improvement in generation was achieved by the Dice metric, with a score of 85.32%, in comparison to Voxel2Mesh. The results demonstrate the potential for accurate lumen geometry generation, with metrics approaching those of the segmentation task. This eliminates the necessity for post-processing steps typically required for the latter. CONCLUSION: Shows promising results for accurately generating lumen geometry with performance similar to the segmentation task, eliminating the need for post-processing steps required for the latter.