Click here to

Session: Imaging: CBCT Image Processing, Image Analysis, and Applications [Return to Session]

ROI Based Deep Learning Enhancement for CBCT Based Radiomics Analysis

M Huang1*, Z Zhang2, Y Lai3, Y Chang4, Z Jiang5, J Lee6, F Yin7, L Ren8, (1) Duke University, Department of Radiation Oncology, Durham, NC, (2) Duke University, Durham, NC, (3) the University of Texas at Southwest, TX, (4) Duke University, Durham, NC, (5) Duke University, Durham, NC, (6)Duke University, Department of Radiation Oncology, Durham, NC,(7) Duke University, Department of Radiation Oncology, Durham, NC,(8) Duke University, Department of Radiation Oncology, Durham, NC,

Presentations

SU-IePD-TRACK 1-6 (Sunday, 7/25/2021) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]

Purpose: To investigate the impact of CBCT image quality and ROI-based deep learning image enhancement on radiomics analysis.

Methods: CBCT image quality is affected by scattering, beam hardening and noise, which directly impact its accuracy for radiomics analysis. This study focuses on investigating the accuracy of CBCT-based radiomics analysis and the efficacy of using deep learning to improve the accuracy. Around 900 CBCT projections were simulated from each patient's CT images using Monte Carlo method to contain primary signal, scatter signal and noise. The projections were then fed into the Varian Itool package to mimic clinical CBCT reconstruction. A Pix2Pix Generative-Adversarial-Network was used to improve the CBCT image quality based on CT images as ground truth. The deep learning model was trained for both enhancing the whole body and the ROI around the tumor using 11 lung patients and tested using 3 patients. Radiomics analysis was performed for both CBCT and enhanced CBCT and compared to ground-truth CT to evaluate their accuracy.

Results: The CBCT image quality was significantly improved after the AI-based enhancement. 18 First order, 22 GLCM, 16 GLRLM, 16 GLSZM, 14 GLDM, and 688 wavelet features were computed for the CT, original CBCT, whole body-based enhanced CBCT, and ROI-enhanced CBCT. Under best scenario, the first 86 features improvement was 83.4%, 18.9%, 13.3%, -11.9% and 17.6% for each radiomics feature zone based on ROI image augmentation, and improved 88.9%, 21.7%, 15.9%, -12.2%, 18.2% based on whole-body image augmentation. The sensitivity of features against image quality was consistent for all 3 patients.

Conclusion: Image quality is vital for establishing accurate CBCT based radiomics analysis, and deep learning can effectively improve the radiomic feature accuracy, especially for low-order features. The superior performance for low-order features is likely due to their consistency with the L1-loss function used in deep learning training.

Funding Support, Disclosures, and Conflict of Interest: funding sources NIH 1R01EB02834-01 and R01CA184173

ePosters

    Keywords

    Not Applicable / None Entered.

    Taxonomy

    IM/TH- Cone Beam CT: Radiomics

    Contact Email

    Share: