SynthPET - FDG-PET synthesis from T1w MRI and ASL CBF for AD
Synthetic FDG-PET can be generated with strong predictive performance using mostly T1w atrophy, supplemented by metabolism-coupled ASL CBF.
FDG-PET is costly and introduces radiation exposure through tracers. Using primarily ADNI data, SynthPET was created to leverage common MRI sequences such as T1-weighted (T1w) and ASL CBF (arterial spin labeling, cerebral blood flow) to create synthetic FDG-PET images. T1w is essential to structural and anatomical brain understanding, while ASL CBF provides a complementary understanding of blood flow/perfusion signal coupled to glucose metabolism, which is also present on FDG-PET. Architecture was a custom cGAN, trained on 1170 T1w patients and fine-tuned on 220 paired T1w+ASL patients. PENN ADRC/FTD data were evaluated for generalization. Validation tasks: 1. Differentiate CN/MCI/AD patients, where synthetic FDG-PET outperformed even GT (before FDR correction), 2. Stats and ablations which showed T1w did the heavy lifting while ASL and transfer learning contributed modestly, 3. FDG-PET signal in clinically relevant brain regions where T1w+ASL correlated with GT in 34/34 regions, 4. LMM showing synthetic FDG-PET captured better late-disease-stage differences while ASL showed better early-stage, and 5. Subtyping CN/non-AD/AD dementia, where encoder features from T1w+ASL-generated synthetic FDG-PET performed best.
This is interesting because the cohort was relatively small, GT-evaluation visual classification tasks were difficult and noisy as referenced by the pilot reader experiment’s low GT-based specificity, and cGANs have been, from my anecdotal experience, inferior to diffusion networks in terms of image generation fidelity. I wonder how this would look with a slightly varied preprocessing pipeline, swapping the cGAN to a diffusion-based solution, and increasing the dataset in some way to balance the need for a slightly larger cohort. The best subtyping coming from encoder features suggests the usefulness of a multimodal foundation network in general for brain imaging-related tasks which could also incorporate genetics and clinical variables. Ultimately, this version of synthetic FDG-PET seems like it could very well be mostly a re-rendering of atrophy on T1w, given how much it carried in validation predictive power over ASL-integrated variants.