Harmonizing Inter-Site Differences in T1-Weighted Images Using CycleGAN

Abstract

IntroductionWhen neuroimaging studies using magnetic resonance imaging (MRI) are conducted across multiple centers, they often encounter inter-site differences in MRI equipment and protocols leading to biases and confounding effects in MRI measurements. There are existing techniques for correcting these site effects, i.e., harmonization, but they have limitations, including the need for preprocessing of MRI data, which involves processes such as spatial normalization. Deep learning-based methods have emerged as potential alternatives that can handle site effects without the need for preprocessing steps. In this study, we propose a novel method based on the generative adversarial network (GAN) framework, CycleGAN, that effectively addresses inter-site differences in T1-weighted images with minimal preprocessing requirements. We compare the harmonization efficacy of CycleGAN with that of the commonly used method ComBat. MethodsWe trained the proposed CycleGAN method and the comparative ComBat method using data from 40 subjects at each of two sites. To evaluate the effectiveness of the two methods, we used data from nine subjects who underwent imaging at both sites. We assessed harmonization performance at the image level using the structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR). Additionally, we evaluated harmonization results at the feature level by analyzing regional cortical thickness and volume data. Cohens d was employed to quantify the differences between feature values. ResultsAt the image level, the ComBat method decreased the median baseline SSIM value from 0.86 (interquartile range [IQR], 0.02) to 0.84 (IQR, 0.02), whereas the proposed CycleGAN method maintained the SSIM value at 0.86 (IQR, 0.02). For PSNR, the baseline value was 18.33 (IQR, 1.78), which decreased to 15.30 (IQR, 2.20) after applying ComBat, but increased to 19.58 (IQR, 3.12) with the proposed CycleGAN method. These findings indicate that CycleGAN preserved the structural and signal similarity of the images. At the feature level, the effect size for cortical thickness decreased from 0.97 (IQR, 1.79) to 0.91 (IQR, 1.54) after applying ComBat, whereas the proposed CycleGAN method yielded an effect size of 1.05 (IQR, 1.14). For cortical volume, the effect size decreased from 0.95 (IQR, 1.78) to 0.69 (IQR, 1.00) after applying ComBat, and decreased to 0.88 (IQR, 0.74) with the CycleGAN method. Compared with baseline, Cohens d was significantly lower with both ComBat (p = 0.000002) and CycleGAN (p = 0.028) with no significant difference between the two methods, indicating similar performance of the two methods under the study conditions. ConclusionThe results underscore the ability of CycleGAN to harmonize data without explicit normalization and emphasize the potential impact of the normalization process on harmonization procedures. Our findings suggest that CycleGAN holds promise as a harmonization technique in multi-site neuroimaging studies.