Generative deep learning model‐based manifold construction for ATN biomarkers

Abstract

AbstractBackgroundGenerative models have the potential to be beneficial in neuroimaging, as they can produce new data samples that closely resemble the ones they were trained on. This can assist in addressing the scarcity of brain scan data, which is a significant issue in neuroscience.MethodIn this study, we developed a one‐ring convolution‐based generative adversarial network (GAN) to generate ATN biomarkers, which operate on an icosahedron cortical mesh structure on a cortical surface. Data from the Alzheimer’s Disease Neuroimaging Initiative cohort was used in the study, including AV45, AV1451, and T1 magnetic resonance imaging (MRI) scans from 2283 subjects. The cortical thickness of all T1 MRI images was calculated using a Freesurfer pipeline and resampled into icosahedron subdivision level 3. AV45 and AV1451 scans were then converted into standard uptake value ratio (SUVR) and mapped to the cortical surface before being resampled into icosahedron subdivision level 3. For each modality, the proposed GAN was trained and evaluated using density and coverage metrics, which measure the fidelity and diversity of generated samples. Also, we validated the efficacy of the constructed ATN manifolds using the data augmentation scenario. Specifically, we evaluated the performance of tau stage classifiers using cortical thickness data with and without augmenting tau PET data from our manifolds.ResultOur proposed GAN achieved high performance in density and coverage. Specifically, the density and coverage for the amyloid SUVR were 1.0 and 0.97, for the tau SUVR were 1.34 and 1.0, and for the cortical thickness were 1.66 and 0.98, respectively, showing high performance in generating realistic and diverse samples for all modalities. The effectiveness of our proposed GAN was further demonstrated to show improved performance in the Tau Braak stage classification experiment when using the generated data. When augmented with generated data, the classification accuracy was improved from 0.77 to 0.84 in terms of AUC scores.ConclusionOur proposed GAN achieved high performance in density and coverage, generating realistic and diverse samples for each ATN biomarker, and demonstrated improved performance in a Tau Braak stage classification experiment when using the generated data.