Author: Trevor McKeown, Deshan Yang, Zhendong Zhang 👨🔬
Affiliation: Duke University, Department of Radiation Oncology, Duke University 🌍
Purpose: Accurate delineation of liver blood vascular structures is crucial for planning and executing therapeutic interventions in liver-related medical procedures. However, current auto-segmentation models often severely under-segment liver vessel trees, primarily due to the limited availability of high-quality, annotated datasets for training the segmentation models. This study presents a novel framework for simulating realistic liver CT images and blood vessel structures. The goal was to provide high-quality image and segmentation ground truth dataset pairs of a virtually unlimited quantity for liver vessel segmentation model development.
Methods: A computational liver vessel generation method was developed to generate high-resolution vessel tree structures. Image voxel intensities for vessels and background tissues were statistically modeled from contrast-enhanced liver CT scans. Image artifacts and imperfections such as quantum noise and motion blurring were digitally simulated via simulated CT acquisition and reconstruction, capturing diverse vascular complexities and realistic image qualities. Using these simulated datasets, a 3D nnUNet model was trained using a binary cross-entropy plus Dice loss to address class imbalance. Performance was evaluated against models trained on two widely used public datasets using an additional 7 manually annotated high-quality ground truth datasets.
Results: unUNet models trained on our simulated dataset demonstrated Dice scores of 0.78 ± 0.04, 24% higher than the best dice scores obtained by nnUNet models trained on the public datasets.
Conclusion: This study demonstrated that high-quality liver vessel segmentation ground truth datasets can be digitally synthesized to address the scarcity of high-quality data. Robust liver vessel segmentation models could be successfully trained on synthesized datasets and obtain improved performance for supporting therapeutic interventions.