Comparison of Patient-Specific Deep Learning Models for Enhancing 4D-CBCT Image for Radiomics Analysis
Z Zhang1*, M Huang2, Z Jiang3, Y Chang4, K Lu5, Z Qiao6, F Yin7, P Tran8, D Wu9, C Beltran10, L Ren11, (1) Duke University, Durham, NC, (2) Mayo Clinic Florida, Jacksonville, FL, (3) Nanjing University, Nanjing, 32, CN, (4) University of Hospital of Pennsylvania, Philadelphia, PA, (5) Duke University, Chapel Hill, NC, (6) University of Florida, Gainesville, FL, (7) Duke University Medical Center, Chapel Hill, NC, (8) University of Maryland, Baltimore, MD, (9) University of Florida, ,,(10) Mayo Clinic, Jacksonville, FL, (11) University of Maryland, Baltimore, MD
Presentations
PO-GePV-I-13 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: Radiomics features extracted from 4D-CBCT provide valuable information for outcome prediction throughout the treatment courses. However, due to the time and dose constraint, 4D-CBCT images suffer from severe streak artifacts caused by under-sampling, which can affect the accuracy of radiomic features. This study aims to investigate patient-specific models to optimize the 4D-CBCT image quality for individual patients for radiomic analysis.
Methods: Different from group-based models, patient-specific models use training data from only one patient to optimize the model. In this study, patients with multiple 4D-CT scans were accrued. 4D-CBCT was simulated from the first 4D-CT by FDK algorithm and then augmented with translation, rotation, and deformation. This augmented 4D-CBCT dataset was used to train a patient-specific model developed based on Pix2pix network to enhance 4D-CBCT to match with ground-truth 4D-CT. The model was then tested to enhance 4D-CBCT simulated from the second 4D-CT. The accuracy of the enhanced 4D-CBCT was evaluated by comparing its radiomic features with ground-truth 4D-CT features. The impact of model dimensionality, region of interest (ROI) selection, and loss function were also investigated by modifying the structure and loss functions of the model.
Results: Patient-specific models with different dimensionality achieved comparable results for improving the radiomics accuracy. Using whole-body or whole-body+ROI L1 loss for the model achieved better results than using the ROI L1 loss alone. For example, the LLL max feature improved 44.5% on average by replacing ROI loss with WB+ROI loss. Overall, patient-specific training models could further improve the accuracy of radiomic features compared to the group-based model.
Conclusion: This study demonstrated the efficacy of using patient-specific deep learning models to enhance 4D-CBCT to improve its radiomic feature accuracy. ROI selection can impact the accuracy of the enhancement. Patient-specific model opens a new horizon for customized image enhancement to achieve high precision radiomic analysis.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by the National Institutes of Health under Grant No. R01-CA184173 and R01- EB028324.
Keywords
Cone-beam CT, Quantitative Imaging
Taxonomy
Contact Email