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

Abstract

Objective: Generative adversarial network (GAN) based methods for MRI standardization are compared to conventional prepro- cessing and a posteriori methods proposed in the literature.

Approach: T2w MR images from 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 gathered, including 86 and 160 T2w MR images taken before radiotherapy (RT) and brachytherapy (BT), 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 and compared to Histogram-matching standardization, z-score normalization, and ComBat harmo- nization method. Different image quality metrics were extracted from Cohort 1 images and the impact of standardization methods was assessed with principal component analysis (PCA). Using Intra-Class Correlation (ICC) and Concordance Correlation Coefficient (CCC), robust features were characterized (CCC&ICC ≥ 0.75). Different ML models were trained to investigate the impact of these harmonization methods on stage classification and relapse prediction, respectively.

Main results: PCA on quality metrics showed that TR and VS changes were mitigated the most with cGAN. Regarding TR/VS modulation, on 1st and 2nd-order features, cGAN achieved the best results with 18/18 and 58/75 of robust features, respectively. On both clinical tasks, AUC improved after stan- dardization. For tumor stage classification, the application of a CycleGAN strategy significantly improved the performances of the ML models compared to classification using raw images.

Significance: GAN-based standardization in MRI might be an additional building block for robust radiomic signatures at a multicentre scale. .