Accelerating cardiovascular model building with convolutional neural networks

The objective of this work is to reduce the user effort required for 2D segmentation when building patient-specific cardiovascular models using the SimVascular cardiovascular modeling software package. The proposed method uses a fully convolutional neural network (FCNN) to generate 2D cardiovascular segmentations. Given vessel pathlines, the neural network generates 2D vessel enhancement images along the pathlines. Thereafter, vessel segmentations are extracted using the marching-squares algorithm, which are then used to construct 3D cardiovascular models. The neural network is trained using a novel loss function, tailored for partially labeled segmentation data. An automated quality control method is also developed, allowing promising segmentations to be selected. Compared with a threshold and level set algorithm, the FCNN method improved 2D segmentation accuracy across several metrics. The proposed quality control approach further improved the average DICE score by 25.8%. In tests with users of SimVascular, when using quality control, users accepted 80% of segmentations produced by the best performing FCNN. The FCNN cardiovascular model building method reduces the amount of manual segmentation effort required for patient-specific model construction, by as much as 73%. This leads to reduced turnaround time for cardiovascular simulations. While the method was used for cardiovascular model building, it is applicable to general tubular structures. Graphical Abstract Proposed FCNN-based cardiovascular model building pipeline. a.) Image data and vessel pathline supplied by the user. b.) Path information is used to extract image pixel intensities in plane perpendicular to the vessel path. c.) 2D images extracted along vessel pathlines are input to the FCNN. d.) FCNN acts on the input images to compute local vessel enhancement images. e.) Vessel enhancement images computed by the FCNN, the pixel values are between 0 and 1 indicating vessel tissue likelihood. f.) The marching-squares algorithm is appliedto each enhanced image to extract the central vessel segmentation. g.) 2D extracted vessel surface points overlayed on original input images. h.) The 2D vessel surface points are transformed back to 3D space. i.) 3D crosssectional vessel surfaces are interpolated along the pathline to form the final vessel model.