(Tuesday, 7/14/2020) 1:00 PM - 2:00 PM [Eastern Time (GMT-4)]
Room: Track 1
Purpose: the impact of 4D-CBCT image quality on radiomics analysis.
Methods: patients’ 4DCT data obtained from clinical patients has been used for FDK reconstruction, training, testing and evaluation. 4D CBCT data were simulated from DRRs of 4D CT using FDK reconstruction with 120 projections. A deep learning model (TecoGAN) was trained on 6 patients to augment the 4D CBCT images quality to match with the corresponding ground-truth 4DCT. The model takes 10 phases of CBCT as input and generates 10 phases of augmented CBCT. 3 patients were used for testing to augment the 4D CBCT images. The augmented images were compared with original 4DCT and original 4D CBCT for radiomic features extraction. In features analysis, four groups of radiomics, including basic, gray-level, texture and wavelet, were evaluated. Delta radiomics between 4D-CBCT before and after augmentation were also calculated.
Results: anatomical details and edge information have been enhanced and artifacts have been removed after augmentation. Based on the basic radiomics evaluation, which include variance, mean and kurtosis, the augmented CBCT reduced the radiomics calculation errors by 51.1% compared to the original CBCT. Specifically, both of the gray-level dependence matrix and texture radiomics reduced the errors by 43% and 75.2% separately. Low level basic features are more sensitive against image quality than high level features. The sensitivity of radiomics features are consistent across different respiratory phases. The findings were consistent across the three patients studied.
Conclusion: study demonstrated that 4D-CBCT image quality has a significant effect on the radiomics analysis. Cautions need to be taken to ensure adequate image quality before performing any radiomics study. The deep learning based augmentation technique also proved to be an effective approach to enhance the image quality for the analysis.
Funding Support, Disclosures, and Conflict of Interest: This work is supported by the National Institutes of Health under Grant No. R01-CA184173 and R01-EB028324.
Keywords
Texture Analysis, Cone-beam CT
Taxonomy
IM- Cone Beam CT: Machine learning, computer vision
Contact Email
