Fine-Tuning Cycle-GAN for Domain Adaptation of MRI Images

TL;DR

The paper tackles MRI domain shifts across scanners by proposing an unsupervised domain adaptation framework based on a fine-tuned Cycle-GAN to translate between T1 and T2 modalities while preserving anatomical structures via cycle-consistency and disparity losses. The model uses two generators and two discriminators with a combined objective that includes adversarial, cycle-consistency, and disparity terms, optimized with a staged training protocol. Evaluation on the iSeg/ISeg pediatric brain MRI dataset shows that the fine-tuned Cycle-GAN achieves superior structural fidelity and cross-domain alignment, evidenced by lower Bhattacharyya distance and higher histogram correlation, as well as higher SSIM after translation. The approach demonstrates robust, bidirectional modality translation without paired data, with potential to improve diagnostic accuracy and consistency in multi-site MRI analyses; future work extends to additional modalities.

Abstract

Magnetic Resonance Imaging (MRI) scans acquired from different scanners or institutions often suffer from domain shifts owing to variations in hardware, protocols, and acquisition parameters. This discrepancy degrades the performance of deep learning models trained on source domain data when applied to target domain images. In this study, we propose a Cycle-GAN-based model for unsupervised medical-image domain adaptation. Leveraging CycleGANs, our model learns bidirectional mappings between the source and target domains without paired training data, preserving the anatomical content of the images. By leveraging Cycle-GAN capabilities with content and disparity loss for adaptation tasks, we ensured image-domain adaptation while maintaining image integrity. Several experiments on MRI datasets demonstrated the efficacy of our model in bidirectional domain adaptation without labelled data. Furthermore, research offers promising avenues for improving the diagnostic accuracy of healthcare. The statistical results confirm that our approach improves model performance and reduces domain-related variability, thus contributing to more precise and consistent medical image analysis.

Figures (9)

• Figure 1: a) Architecture of the Proposed GAN. $G_A$ and $G_B$ are denotes the generator network and $D_A$ and $D_A$ are two discriminators, trained by adversarial losses $\mathcal{L}_{GA}$ and $\mathcal{L}_{GB}$, respectively. $x \sim T1$ denotes the image set in the source domain. $y \sim T2$ represents an image set from the target domain. $L_c$ and $L_d$ indicates the cycle consistent and disparity loss, respectively.
• Figure 2: Results show MRI adaptation of T1 modality into T2 by Proposed GAN, CycleGAN, BiGAN, and DualGAN from three different coordinates. Row one two and three shows results from Axial, Coronal, and Sagittal coordinates, respectively.
• Figure 3: Results show MRI adaptation of T2 modality into T1 by Proposed GAN, CycleGAN, BiGAN, and DualGAN from three different coordinates. Row one two and three shows results from Axial, Coronal, and Sagittal coordinates, respectively.
• Figure 4: SSIM map from original and generated T1 and T2 image by BiGAN
• Figure 5: SSIM map from original and generated T1 and T2 image by CycleGAN