Abstract
Synthesizing multi-modal Positron Emission Tomography (PET) images from structural Magnetic Resonance Imaging (MRI) is a significant challenge in Alzheimer’s Disease (AD) research due to the "one-to-many" mapping problem and the need for modality-specific details. This paper introduces the Residual Diffusion Model (ResDM), which leverages residual learning to focus on inter-domain differences and incorporates clinical priors (age, gender) via a text-guided module. Experimental results on the ADNI dataset demonstrate that ResDM outperforms existing state-of-the-art methods in both image reconstruction quality and downstream diagnostic accuracy.
Introduction
Effective diagnosis of Alzheimer’s Disease often relies on ATN analysis (Amyloid, Tau, Neurodegeneration), requiring Aβ-PET, tau-PET, and MRI scans. However, PET imaging is costly and involves radiation exposure. While AI-driven synthesis of PET from MRI offers a solution, traditional models struggle to capture subtle, modality-specific signals and often produce blurry results that lack diagnostic utility.
Methodology
The proposed ResDM framework addresses these issues through three primary components:
- Residual Diffusion Learning: Instead of generating the full PET image, the model learns the residual difference between the MRI and the target PET. This focuses the network on unique pathological signals rather than shared anatomical structures.
- Prior-Information-Guided Module: Clinical metadata (age, gender, modality type) is encoded using a pre-trained CLIP model and injected into the diffusion process as semantic guidance.
- Intra-Domain Difference Loss: A specialized loss function ensures the relationship between synthesized Aβ and tau images mirrors real-world biological discrepancies.
Experimental Results
The model was evaluated using 1,274 image sets from the ADNI dataset. ResDM achieved superior performance in image quality metrics, reaching a PSNR of 26.38 dB for Aβ-PET, significantly outperforming standard diffusion models (DDPM).
Qualitatively, ResDM produces sharper images with lower error rates in the cortex compared to GAN-based or standard diffusion approaches. As shown in the comparison below, the error maps for ResDM are significantly lighter, indicating higher fidelity to the ground truth (GT).
Furthermore, the synthesized images proved highly effective for downstream tasks, achieving 81.90% accuracy for Aβ status classification and 83.02% for tau status classification, closely approaching the performance of real PET scans (94.18% and 92.32%, respectively).
Conclusion
ResDM represents a significant step forward in virtual multi-modal PET synthesis. By combining residual learning with clinical prior guidance, it overcomes the limitations of traditional synthesis models. Future work will focus on improving inference speed and validating the model on broader, multi-center clinical datasets.
Summary updated