(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: investigate the impact of image quality augmentation of CBCT on Radiomics Analysis.
Methods: treated eight SBRT patients images (3-5 fractions) were used as training data. A deep learning model (GAN) was trained based on the above patients to augment CBCT image quality to match the corresponding CT. Three patients were used for testing the effects of image augmentation on radiomics analysis. The augmented CBCT and original CBCT were compared to the Plan CT as baseline. All 547 radiomics of the plan CT and original CBCT and augmented Post-CBCT of the patient were extracted from GTV of each fraction of the patient respectively (7 Histogram, 22 gray level co-occurrence matrix(GLCM),13 gray-level run-length matrix(GLRLM), 13 gray level size-zone matrix(GLSZM), 5 neighborhood gray-tone difference matrix(NGTDM) features, 480 wavelet-based features and seven shape features). All radiomics were evaluated from CBCT to CT, and Post-CBCT to CT; the fractional radiomics feature of Post-CBCT to CBCT comparison were also evaluated.
Results: interclass correlation were calculated for each zone of Radiomics features of each fraction and for each patient, and all features for comparison. The cross correlation between CBCT and CT features improved from 0.896, 0.810, 0.901 in original CBCT to 0.899, 0.974, 0.961, respectively in augmented CBCT for the three patients. In all radiomics zones, especially in low-order features, the improvement of image quality improves the correlation of the CBCT images to Plan CT, which used widely for Radiomics analysis. During all fractions, the correlation of the Augmented CBCT to original CBCT was consistent and were all above 0.9. The sensitivity of features against image quality was consistent among different days’ CBCT. The same phenomenon was observed across all three patients.
Conclusion: quality can significantly impact radiomics analysis using CBCT. Image quality augmentation is vital for establishing accurate and robust CBCT based radiomics related analysis.
Funding Support, Disclosures, and Conflict of Interest: acknowledge the NIH grants NIH R01-CA184173 and R01-EB028324
Keywords
Not Applicable / None Entered.
Taxonomy
IM- Cone Beam CT: Machine learning, computer vision
Contact Email
