SRU-Pix2Pix: A Fusion-Driven Generator Network for Medical Image Translation with Few-Shot Learning

TL;DR

SRU-Pix2Pix addresses the challenge of translating MRI modalities under few-shot conditions by augmenting Pix2Pix with a SEResNet encoder, a U-Net++ decoder, and a PatchGAN discriminator. It employs a $2.5$D input strategy and a composite loss to balance global realism with local structural fidelity, achieving robust performance across BraTS 2023, IXI, and BraTS 2019 without substantial data requirements. The approach delivers consistent improvements in PSNR, SSIM, MS-SSIM, LPIPS, MSE, and NMSE, demonstrating strong generalization and potential clinical applicability for augmenting multimodal MRI datasets. These results position SRU-Pix2Pix as a practical extension of Pix2Pix for reliable, high-fidelity medical image translation under real-world data constraints.

Abstract

Magnetic Resonance Imaging (MRI) provides detailed tissue information, but its clinical application is limited by long acquisition time, high cost, and restricted resolution. Image translation has recently gained attention as a strategy to address these limitations. Although Pix2Pix has been widely applied in medical image translation, its potential has not been fully explored. In this study, we propose an enhanced Pix2Pix framework that integrates Squeeze-and-Excitation Residual Networks (SEResNet) and U-Net++ to improve image generation quality and structural fidelity. SEResNet strengthens critical feature representation through channel attention, while U-Net++ enhances multi-scale feature fusion. A simplified PatchGAN discriminator further stabilizes training and refines local anatomical realism. Experimental results demonstrate that under few-shot conditions with fewer than 500 images, the proposed method achieves consistent structural fidelity and superior image quality across multiple intra-modality MRI translation tasks, showing strong generalization ability. These results suggest an effective extension of Pix2Pix for medical image translation.

Figures (9)

• Figure 1: Representative challenges in medical image translation and the corresponding solutions. (a) GAN: unstable training. (b) VAE: blurred outputs caused by sampling. (c) Diffusion Model: misaligned denoising introduces bias. (d) Diffusion-Zero: image-level focus limits generalization. (e) Our: stable training, high-quality synthesis, and strong generalization.
• Figure 2: An End-to-End Framework for Medical Image Modality Translation: (a) 3D-to-2.5D Data Preprocessing, (b) GAN-Based Generator with SEResNet and U-Net++ and PatchGAN Discriminator, and (c) Encoder Layer Data Flow Details
• Figure 3: Radar charts comparing the performance of multiple models on various MRI image translation tasks across five datasets. Each chart shows seven different models evaluated on multiple metrics (PSNR, SSIM, LPIPS, MS-SSIM , MSE , NMSE). The top row presents three charts corresponding to the BraTS 2023 dataset for T1→T2, T1→FLAIR, and T2→FLAIR tasks. The bottom row shows two charts for the IXI dataset (PD→T2) and the BraTS 2019 zero-shot generalization task. Metrics are normalized for visualization to facilitate a clear comparison of each model's strengths and differences.
• Figure 4: Ablation study on different model configurations for the T1→T2 MRI translation task.(a) Normalized grouped bar chart showing each model's performance across multiple metrics, with values scaled to [0,1] for comparability.(b) Normalized heatmap illustrating relative performance of different models, where color intensity represents the normalized metric values.These visualizations provide complementary insights: the bar chart emphasizes individual metric differences, while the heatmap highlights overall trends and patterns.
• Figure 5: Error heatmaps illustrating the discrepancies between the outputs of quantitative comparison methods and the target images. Darker regions represent smaller errors, while brighter regions represent larger errors.