Basic Science and Pathogenesis.

Area of Science:

• Neuroimaging and computational pathology.
• Application of machine learning in neurology.

Background:

• Accurate dementia diagnosis relies on post-mortem neuropathology, limiting in-vivo assessment.
• Positron emission tomography (PET) offers a potential in-vivo method for predicting neuropathology.
• This research proposes a semi-supervised learning (SSL) model for neuropathology prediction using FDG-PET images.

Purpose of the Study:

• To develop and validate a semi-supervised learning (SSL) model for predicting neuropathology from single FDG-PET images.
• To leverage a patient similarity network (PSN) for label propagation in predicting neuropathological conditions.
• To assess the model's accuracy in differentiating various dementia-related pathologies.

Main Methods:

• A graph-based SSL model was constructed using a patient similarity network (PSN) from 1,495 FDG-PET images.
• Principal component analysis and cosine similarity were used to define image similarity and build the PSN.
• The Poisson Learning Method propagated neuropathologic labels from 204 annotated cases to unlabeled data within the PSN structure.
• Leave-one-out cross-validation (LOOCV) was employed for performance evaluation.

Main Results:

• The model demonstrated high accuracy in predicting Alzheimer's disease, frontotemporal lobar degeneration with tau inclusions, and Lewy body disease.
• A confusion matrix analysis confirmed the predictive performance.
• Sankey diagrams and label distribution analysis showed strong correlations between clinical diagnoses and predicted neuropathology.

Conclusions:

• Graph-based SSL effectively predicts neuropathological changes using FDG-PET imaging, even with limited neuropathologic labels.
• SSL models are valuable for integrating large neuroimaging datasets with scarce neuropathologic data.
• This approach holds promise for advancing dementia diagnosis and understanding.