Ambient-Pix2PixGAN for Translating Medical Images from Noisy Data

TL;DR

The paper addresses translating medical images when measurements are noisy by proposing Ambient-Pix2PixGAN, which fuses Pix2PixGAN with AmbientGAN-style measurement modeling. The approach trains a conditional GAN on degraded measurements by enforcing consistency between generated images and their simulated measurements, using a combined loss $\mathcal{L}=\mathcal{L}_{cGAN}+\lambda\mathcal{L}_{L1}$. Empirical results on MR-to-PET translation show that Ambient-Pix2PixGAN improves traditional image quality metrics and aligns task-based performance with ground truth, indicating robust translation under noise. This method offers a path to reliable cross-modality synthesis in noisy medical imaging and supports stochastic object modeling for objective image quality assessment.

Abstract

Image-to-image translation is a common task in computer vision and has been rapidly increasing the impact on the field of medical imaging. Deep learning-based methods that employ conditional generative adversarial networks (cGANs), such as Pix2PixGAN, have been extensively explored to perform image-to-image translation tasks. However, when noisy medical image data are considered, such methods cannot be directly applied to produce clean images. Recently, an augmented GAN architecture named AmbientGAN has been proposed that can be trained on noisy measurement data to synthesize high-quality clean medical images. Inspired by AmbientGAN, in this work, we propose a new cGAN architecture, Ambient-Pix2PixGAN, for performing medical image-to-image translation tasks by use of noisy measurement data. Numerical studies that consider MRI-to-PET translation are conducted. Both traditional image quality metrics and task-based image quality metrics are employed to assess the proposed Ambient-Pix2PixGAN. It is demonstrated that our proposed Ambient-Pix2PixGAN can be successfully trained on noisy measurement data to produce high-quality translated images in target imaging modality.

Figures (5)

• Figure 1: An illustration of the Ambient-Pix2PixGAN architecture in the image-to-image translation task from MRI ($x$) to PET ($y$). The generator $G$ is trained to generate a fake image $y_c$ that is processed by the image degradation function $\mathcal{H}_n$ to generate the fake measurement data $\hat{y}$ to fool $D$. The discriminator $D$ is trained to classify between fake (synthesized by the generator) and real tuples $\{x, \hat{y}\}$ and $\{x, y\}$.
• Figure 2: (a1) and (a2) are the noisy MRI images as input, (b1) and (b2) are the corresponding PET images generated by our proposed Ambient-Pix2PixGAN.
• Figure 3: (a) and (b) are images generated by the Pix2PixGAN and Ambient-Pix2PixGAN, respectively.
• Figure a1: (a) Singular value spectra comparison, (b) Radially averaged power spectra comparison, (c)-(e) represent probability density functions (PDFs) corresponding to SSIM, PSNR, and RMSE, respectively.
• Figure a1: Comparison between Pix2PixGAN generated, Ambient-Pix2PixGAN generated and ground truth $\text{SNR}_{HO}$ for different signal detection tasks.