GAN-based standardization of MR images: a promising approach for the development of multicentre radiomic models.

Objective.This study evaluated generative adversarial network (GAN)-based magnetic resonance imaging (MRI) standardization methods by comparing them with conventional preprocessing anda posterioriapproaches from the literature in their ability to mitigate the influence of acquisition parameters on radiomic analyses.Approach.MR T2-weighted images (T2w) of 30 patients with locally advanced cervical cancer (LACC) were acquired prospectively (cohort 1). For each patient, three images were taken sequentially on the same scanner with different values of repetition time (TR) and voxel size (VS). A retrospective cohort of 160 LACC patients (cohort 2) was also collected, including 86 and 160 T2w MR images taken before radiation therapy and brachytherapy, respectively. A conditional GAN (cGAN) and a cycleGAN were trained on cohort 1 and cohort 2, respectively, to generate images robust to the impact of acquisition parameters. This generative network-based standardization approach was compared to histogram-matching standardization,z-score standardization, and the ComBat harmonization methods. In this aim, different image quality metrics were extracted from cohort 1 images and the impact of standardization methods was assessed using principal component analysis (PCA). Using intra-class correlation (ICC) and concordance correlation coefficient (CCC), robust features were characterized (CCC&ICC⩾0.75). Different classical ML models were finally trained to investigate the impact of these harmonization methods on stage classification and relapse prediction, respectively.Main results.PCA on quality metrics showed that the changes in TR and VS were mitigated the most with cGAN. Regarding TR/VS modulation, cGAN achieved the best results for the first- and second-order features, with 18/18 and 58/75 robust features, respectively. In both clinical tasks, ROC-AUC improved after standardization. For tumor stage classification, the application of a cycleGAN strategy significantly improved the performance of trained models compared to classification using raw images (ROC-AUC of0.68±0.16before standardization and0.88±0.09after standardization for the best ML model, i.e. logistic regression).Significance.GAN-based standardization in MRI could be an additional building block for robust radiomic signatures on a multicenter scale.