2.5D Multi-view Averaging Diffusion Model for 3D Medical Image Translation: Application to Low-count PET Reconstruction with CT-less Attenuation Correction

TL;DR

The paper tackles reducing radiation exposure in PET by enabling CT-less attenuation correction and directly translating low-count NAC-PET to AC-PET in 3D. It introduces the 2.5D Multi-view Averaging Diffusion Model (MADM), which uses separate 2.5D diffusion models for axial, coronal, and sagittal views, averaged at each denoising step, and leverages a CNN-based prior to accelerate inference. MADM demonstrates superior performance over CNN and diffusion baselines in 3D PET translation under ultra-low-dose conditions, with improvements in both global image quality and lesion quantification. The approach offers a memory-efficient, scalable pathway toward CT-free attenuation correction and dose reduction with potential applicability to other 3D medical imaging tasks.

Abstract

Positron Emission Tomography (PET) is an important clinical imaging tool but inevitably introduces radiation hazards to patients and healthcare providers. Reducing the tracer injection dose and eliminating the CT acquisition for attenuation correction can reduce the overall radiation dose, but often results in PET with high noise and bias. Thus, it is desirable to develop 3D methods to translate the non-attenuation-corrected low-dose PET (NAC-LDPET) into attenuation-corrected standard-dose PET (AC-SDPET). Recently, diffusion models have emerged as a new state-of-the-art deep learning method for image-to-image translation, better than traditional CNN-based methods. However, due to the high computation cost and memory burden, it is largely limited to 2D applications. To address these challenges, we developed a novel 2.5D Multi-view Averaging Diffusion Model (MADM) for 3D image-to-image translation with application on NAC-LDPET to AC-SDPET translation. Specifically, MADM employs separate diffusion models for axial, coronal, and sagittal views, whose outputs are averaged in each sampling step to ensure the 3D generation quality from multiple views. To accelerate the 3D sampling process, we also proposed a strategy to use the CNN-based 3D generation as a prior for the diffusion model. Our experimental results on human patient studies suggested that MADM can generate high-quality 3D translation images, outperforming previous CNN-based and Diffusion-based baseline methods.

Figures (7)

• Figure 1: Illustration of low-count/dose PET reconstruction with CT-less Attenuation Correction (AC) for reducing the overall radiation dose in PET.
• Figure 2: The overall workflow of 2.5D Multi-view Averaging Diffusion Model (MADM). MADM contains a one-step inference generative model(orange) and MAD-BLK(grey). The MAD-BLK contains Three models in axial, sagittal, and coronal view, and the output of each model will be averaged in the average block before output.
• Figure 3: Visual comparison of AC-SDPET generation from different methods under 5% NAC-LDPET settings. The coronal view image(top) and error map(bottom) are shown. RMSE and PSNR values are calculated for each individual volume. The CT-based AC-SDPET(top) and image index in to the color map of the error map(bottom) are shown in the first column.
• Figure 4: Visual comparison of AC-SDPET generation from different methods under 10% NAC-LDPET settings. The coronal view image(top) and error map(bottom) are shown. RMSE and PSNR values are calculated for each individual volume. The CT-based AC-SDPET(top) and image index in to the color map of the error map(bottom) are shown in the first column.
• Figure 5: Visual comparison of AC-SDPET generation form 2D MADM and 2.5D MADM under 5% NAC-LDPET settings. The coronal view image(top) and error map(bottom) are shown. RMSE and PSNR values are calculated for each volume. The CT-based AC-SDPET(top) and image index into the color map of the error map(bottom) are shown in the first column.