Amyloid β pathology prediction in neurodegenerative diseases: deep-learning and early-phase PET approaches

Abstract

Neurodegenerative diseases are increasing as the society ages. The development of positron emission tomography and corresponding radiotracers permits the assessment of various neuropathologic proteins. Previously, single neuropathologic protein were considered to be associated with single disease, but a recent theory of mixed brain pathology proposes that different neuropathologic proteins are distributed throughout the brain and cause various clinical symptoms depending on their distribution and concentration. Amyloid beta is one such neuropathologic protein, and is closely associated with Alzheimer's disease. Amyloid beta has been found in numerous other neurodegenerative disorders with the advancement of brain imaging technology, and its function has been researched. Despite the clinical significance, brain imaging using different radioactive trackers is challenging in a single disease because of issues like cost and radiation exposure.

In this study, I used fluorodeoxyglucose positron emission tomography and deep-learning (DL) techniques to build a model for predicting amyloid beta pathology in Alzheimer's disease groups, and I tested it using a second dataset. Next, I determined the relationship between early-phase positron emission tomography and amyloid beta pathology in the Parkinson's disease and dementia with Lewy-bodies groups. Subsequently, I created a model to predict amyloid beta pathology using early-phase positron emission tomography. In addition, I investigated whether brain perfusion differs according to the presence of amyloid beta neuropathologic protein in the Parkinson's disease and dementia with Lewy-bodies groups.

First, I confirmed the amyloid beta-positive/negative regions using fluorodeoxyglucose positron emission tomography. Then, using DL, I developed a prediction model for amyloid beta-positive/negative in Alzheimer's disease that outperformed the model used in previous studies. The model was confirmed to learn from the known data when it was validated using the novel method I had suggested. Next, using early-phase positron emission tomography, I determined the association between areas with altered perfusion in the Parkinson's and dementia with Lewy bodies groups with the presence or absence of amyloid beta. I developed a model to predict amyloid beta pathology using this perfusion data and logistic regression models. This model performed similarly to one I had used in an earlier study on Alzheimer's disease. Additionally, in the groups with Parkinson's disease and dementia with Lewy bodies, brain perfusion sites differed depending on the presence or absence of amyloid beta neuropathologic proteins.

In conclusion, I used fluorodeoxyglucose positron emission tomography, DL, and early-phase positron emission tomography to create amyloid beta pathology prediction models in various neurodegenerative diseases. These results reveal the potential of amyloid beta deposition prediction models in different neurodegenerative disorders and demonstrate how different radiotracers can provide molecular data.