U-net for automated thoracic CT semantic segmentation

Cardiac computed tomography has a clear clinical role in the evaluation of coronary artery disease and assessment of coronary artery calcium (CAC) but the use of ionizing radiation limits the clinical use. Beam-shaping “bow-tie” filters determine the radiation dose and the effective scan field-of-view diameter (SFOV) by delivering higher X-ray fluence to a region centered at the isocenter. A method for positioning the heart near the isocenter could enable reduced SFOV imaging and reduce dose in cardiac scans. We developed a predictive approach to center the heart and reduce the SFOV. As part of this effort, we used a UNet to segment noncontrast thoracic CT scans to estimate the associated dose reductions. Here we publish the UNet network.
Specifically, this dataset contains a trained U-net (convolutional neural network) which was trained for the purpose of segmenting noncontrast thoracic computed tomography images. , We collected noncontrast thoracic CT images from our institution and manually segmented them. We then trained a U-Net (with the Pytorch framework) to perform semantic segmentation. The final state of the trained network is contained in this dataset., This repository contains a .pth file which is the complete set of trained weights for the neural network. A repository of Python code contained at https://github.com/ucsd-fcrl/unet_deploy may be a useful starting point for using this U-Net.