Cross-Dataset Multiple Organ Segmentation from CT Imagery Using FBP-Derived Domain Adaptation

Multi-organ segmentation from whole-body computed tomography (CT) scans has gained increasing research interest over recent years. While the learning-based segmentation algorithm has lately achieved tremendous success, the need for detailed annotation of multiple organs further increases the manual burden. With a limited number of annotated volumetric datasets, it would be beneficial to apply the trained model from such a set to CT images acquired from other sites with different scanners. Nevertheless, the discrepancy among training and testing images significantly deteriorates segmentation performance. While there are many domain adaptation efforts, in this work we proposed a filtered back-projection based algorithm for performing domain adaptation for CT imagery. An optimal CT reconstruction kernel was obtained by minimizing the disparity between two images. Furthermore, since the Gaussian kernel is an eigen-function of the Fourier transformation, the adaptation computation was proven to be simple linear filtering. The proposed method was tested and compared with multiple methods to demonstrate improvement by employing such a model/theory-based adaptation approach. The proposed method, used in conjunction with a common convolutional neural network, such as the U-Net or V-Net, with or without the domain adaptation, achieves high accuracy in a multiple-organ segmentation task. Approximately 30% of data was used for training, 70% was used for testing, and an average dice of 0.88 was achieved in 8 organs.