Development of a Radiomics-Base Glioblastoma Prognosis Prediction Model Using CycleGAN
H Yoshimura1, 2*, D Kawahara2, (1) Kure medical center, Hiroshima, 34, JP, (2) Hiroshima university, Hiroshima, 34, JP
Presentations
PO-GePV-M-289 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: To propose the style transfer model for multi-contrast MRI images with Cycle-consistent Generative adversarial network (CycleGAN), and evaluate image quality and prognosis prediction performance.
Methods: Style transfer models of T1 weighted MRI image (T1w) to T2 weighted MRI image (T2w) and T2w to T1w with CycleGAN were constructed with BraTS dataset. The style transfer model was validated with the TCGA-GBM dataset. Moreover, imaging features were extracted from real and synthetized images. These features were transformed to rad-scores by least absolute shrinkage and selection operator (LASSO)-Cox regression. The prognosis performance was estimated by Kaplan-Meier method.
Results: The survival time has a significant difference between good and poor prognosis groups for both of real and synthesized T2w (p<0.05). However, it had no significant difference for both of real and synthesized T1w. On the other hand, there was no significant difference of the real and synthesized T2w in both of good and poor prognosis. The results of T1w were similar in the point that there was no significant difference of the real and synthesized T1w. It was suggested that the synthetized image is equivalent to the real image in comparison with the imaging features. It was found that the synthetized image could be used for prognosis prediction.
Conclusion: The proposed prognostic model using CycleGAN could reduce the cost and time of image scanning, which lead to promote building the patients outcome prediction with multi-contrast images.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email