Automatic generation of training data from dual-energy CT scans: a new approach to train renal vessel segmentation models.

Authors

Abstract

To investigate the feasibility of automatic training data generation using virtual non-contrast (VNC) images of dual-energy CT and to evaluate the performance of medical segmentation models trained on auto-generated data. Training images were generated from contrast-enhanced CT (CECT) scans using VNC algorithm. Corresponding training labels were generated from iodine-density images via intensity thresholding. The paired VNC images and labels constituted the dataset (791 cases in total) for training non-enhanced CT (NECT) segmentation models. We attempted to train segmentation models with VNC images alone, as well as the combination of VNC and real NECT images. Models trained with auto-generated data effectively performed renal vein segmentation on NECT images, achieving Dice score of 69.60%. It identified 100% (70/70) vein main trunks and 54.5% (6/11) accessory veins in the test set. The models' performance on artery segmentation was not satisfactory, with Dice score of 44.56%. It only identified 41.4% (29/70) artery main trunks and 4.54% (1/22) accessory arteries. However, the combination of generated data and labeled data greatly improved model's artery Dice to 71.23%. Combined model identified 100% (70/70) vein main trunks, 98.6% (69/70) artery main trunks, 68.2% (15/22) accessory arteries and 63.6% (7/11) accessory veins. Its accessory artery detection is significantly higher than the model trained with labeled data (15/22 vs. 5/22, p < 0.01), closing the gap with human radiologist (15/22 vs. 18/22, p = 0.250). VNC and Iodine-density images can be utilized to automatically generate training data for segmentation models. This pipeline shows promise for training segmentation models on NECT images.

Tags

Topics