MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Deep-Learning-Based CT ImageStandardization to Improve Stability of Radiomics Features in Non-Small CellLung Cancer

J Zhang1*, M Selim2, M Brooks3, B Fei4, G Zhang5, J Chen6, (1) University of Kentucky, Lexington, KY, (2) University Of Kentucky, ,,(3) University Of Kentucky, ,,(4) University of Texas (UT) at Dallas and UT Southwestern Medical Center, Richardson, TX, (5) Uthealth, ,,(6) University Of Kentucky,

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: Radiomics has shown promising in predicting prognosis and therapeutic response. However, the discrepancy in the underlying image data due to the use of various acquisition parameters creates a limit for large-scale cross-center studies. This study is to develop an end-to-end CT image standardization approach to improve the stability of radiomics features in Non-Small Cell Lung Cancer (NSCLC).

Methods: A deep learning approach named STAN-CT based on GAN models was developed. It adopted a novel one-to-one mapping loss function on the latent space, enforcing the generator to draw sample distribution from the same distribution where the target image belongs to. We used a feature-based loss to improve the performance of the discriminator. Also, STAN-CT introduced a DICOM reconstruction framework that integrated all the synthesized image patches to generate a DICOM for clinical use. Three quality control units were inserted into the framework, each being adopted to address a specific image quality problem. STAN-CT was trained and tested using images of patients with NSCLC. CT images were acquired in a Siemens Force using three kernels (Bl57, Bl64 and Br40) and different slice thickness. The training data included 14,699 images while the test data contained 3,810 images. Five radiomics features (dissimilarity, contrast, homogeneity, energy, correlation) were evaluated and compared before and after image normalization.

Results: Compared to Bl64, the averaged absolute error of 5 radiomics features before and after normalization reduces from 37.52% to 9.94% for Bl57 and from 43.34% to 21.34% for Br40. The use of DICOM reconstruction framework with quality control units achieves significantly better results than using GAN models directly, indicating the quality control units in our framework are critical regarding artifact detection and removal.

Conclusion: STAN-CT provides an end-to-end solution for CT image standardization and normalization. More radiomics features are to be evaluated to validate the proposed method.

Funding Support, Disclosures, and Conflict of Interest: This study is supported by NIH NCI (no. 1R21CA231911) and Kentucky Lung Cancer Research (no. KLCR-3048113817).

Keywords

CT

Taxonomy

IM/TH- Image Analysis (Single Modality or Multi-Modality): Machine learning

Contact Email