High-Resolution 5D Cardiac Photon-Counting Micro-CT in Mice Using a Fast Turntable Scanner and Deep Learning-Based Full Heart Segmentation
Preclinical 5D (3 spatial dimensions + energy + time) in vivo photon-counting micro-CT (PCCT) imaging enables cardiac phenotyping in humanized mouse models of cardiovascular disease risk. Our group has previously used 5D cardiac PCCT imaging of mice and a Unet convolutional neural network that we trained for multi-chamber heart segmentation to investigate how genetic factors such as apolipoprotein E (APOE) genotype and lifestyle factors such as diet and exercise impact cardiac functional metrics. However, these studies relied on our previous PCCT system that required rotation of the mouse and limited us to an isotropic voxel size of 125 µm that is too large to resolve small substructures or subtle changes in the heart. We recently integrated our photon-counting detector into a new, higher magnification (enabling 75 µm voxel size) micro-CT system that rotates the detector and x-ray source on a turntable while limiting animal motion to vertical translation. This work describes our efforts to transfer our cardiac phenotyping pipeline to the new PCCT system. This includes a new acquisition protocol and transfer of existing intrinsic temporal gating, 5D iterative reconstruction, and material decomposition algorithms. We demonstrate a new dedicated Unet denoising solution that improves computational efficiency and contrast to noise ratio relative to iterative reconstruction. We improved Dice coefficients for Unet multi-chamber heart segmentation by retraining on labeled data from the turntable system. This work demonstrates, for the first time, 5D cardiac phenotyping on a turntable PCCT system, including intrinsic gating, deep-learning-based denoising, and multi-chamber segmentation at 75 µm resolution.
