CT Image Standardization for Radiomic Feature Enhancement in Non-Small Cell Lung Cancer
M Selim1*, J Chen2, B Fei3, G Zhang4, J Zhang5, (1) University Of Kentucky, ,,(2) University Of Kentucky, ,,(3) University of Texas (UT) at Dallas and UT Southwestern Medical Center, Richardson, TX, (4) Uthealth, ,,(5) University of Kentucky, Lexington, KY
Presentations
PO-GePV-I-18 (Sunday, 7/25/2021) [Eastern Time (GMT-4)]
Purpose: This study is to develop an end-to-end fully automatic CT image standardization approach named RadiomicGAN to improve the stability of radiomics features in NSCLC.
Methods: RadiomicGAN was developed based on transfer learning and Generative Adversarial Network (GAN), which can efficiently learn the standard image feature distributions from relatively small medical image data. The performance of RadiomicGAN was enhanced by leveraging the wide dynamic range of CT imaging. Specifically, the model adaptively identified the effective pixel range in the current training cycle and smartly optimized its loss in the next training cycle. CT images for training and testing were acquired in a Siemens Sensation Force using three different kernels (Bl57, Bl64, Br40) and different slice thickness (0.5, 1, 1.5 and 3mm). The training data included a total of 14,372 images, while each test data contained 387 images. The RadiomicGAN was also trained and tested using CT images of a chest phantom. Radiomic features were extracted from the lung nodule delineated manually with the IBEX. In total, 1,401 radiomics features were evaluated and compared before and after image normalization.
Results: RadiomicGAN achieves significantly better image normalization performance than other state-of-art models. In average, using Bl64 as the standard image reconstruction kernel, the averaged absolute error of the 1,401 radiomics features before and after normalization reduced from 51.82% to 29.8% for images reconstructed using Bl57 and from 78.45% to 30.5% for images reconstructed using Br40. The existing CT image standardization models such as Choe et al, GANai and STAN-CT reduced the error to 44.7%, 42.1%, and 40.5% respectively for Bl57, and 56.1%, 54.5%, and 48.6% respectively for Br40.
Conclusion: RadiomicGAN provides a better end-to-end solution for CT image standardization and normalization. Further evaluation with larger amount of patient data are warranted.
Funding Support, Disclosures, and Conflict of Interest: This study is partially supported by NIH
Keywords
CT, Image Analysis, Quantitative Imaging
Taxonomy
IM/TH- Image Analysis (Single Modality or Multi-Modality): Machine learning
Contact Email