Recognizing early signs of dementia utilizing [18F]-fluorodeoxyglucose positron emission tomography brain imaging data and machine learning
Savila, Veikka (2025-06-06)
Recognizing early signs of dementia utilizing [18F]-fluorodeoxyglucose positron emission tomography brain imaging data and machine learning
(06.06.2025)
Julkaisu on tekijänoikeussäännösten alainen. Teosta voi lukea ja tulostaa henkilökohtaista käyttöä varten. Käyttö kaupallisiin tarkoituksiin on kielletty.
avoin
Julkaisun pysyvä osoite on:
https://urn.fi/URN:NBN:fi-fe2025063075964
https://urn.fi/URN:NBN:fi-fe2025063075964
Tiivistelmä
Artificial intelligence, powered by advanced machine learning models, has revolutionized scientific research. In parallel, available neuroimaging data has increased. However, the utilization of machine learning models to analyse the data has been limited to small datasets, producing only directional results. Furthermore, processing of the positron emission tomography (PET) imaging data has varied, generating results inapplicable for cross-comparison. This thesis aimed to provide a reproducible and generalizable analysis of [18F]-Fluorodeoxyglucose (FDG) PET images with focus on revealing underlying changes in brain metabolism preceding dementia. Magia, a standardized pipeline for analysis of the PET scans, was utilized, as well as machine learning models on the data acquired from Magia, for feature extraction and classification between diseased and healthy subjects.
This thesis applied methods from previous studies in the fields of neuroimaging, neuroinformatics, and machine learning. The PET imaging data, with available memory problem diagnoses, was extracted from AIVO database of Turku PET Centre. The Magia pipeline for image analysis utilized three PET image quantification methods: 1) standardized uptake value, 2) fractional uptake rate, and 3) linear regression Patlak plot. The results of the three quantification methods in combination with the diagnostic data from AIVO were used in the training of three separate machine learning models. The machine learning models were used to classify between A) healthy controls, B) other memory problems, C) memory disorders, and D) other subjects.
The produced prediction models were able to outperform an uninformed guess but performed worse than models in prior studies. Achieving robust prediction models proved challenging for example due to misplaced data. Furthermore, due to underrepresentation of subjects within groups during the training of the models, the prediction models exhibited poor sensitivity. Multiple points of improvement were characterized during the study. Further research in the topic is required. A robust prediction model relying on PET image derived data rather than the images themselves will decrease the required resources in the study of memory problems. Over and above that, the features extracted will contribute to a better understanding of the pathophysiology of neurodegeneration.
This thesis applied methods from previous studies in the fields of neuroimaging, neuroinformatics, and machine learning. The PET imaging data, with available memory problem diagnoses, was extracted from AIVO database of Turku PET Centre. The Magia pipeline for image analysis utilized three PET image quantification methods: 1) standardized uptake value, 2) fractional uptake rate, and 3) linear regression Patlak plot. The results of the three quantification methods in combination with the diagnostic data from AIVO were used in the training of three separate machine learning models. The machine learning models were used to classify between A) healthy controls, B) other memory problems, C) memory disorders, and D) other subjects.
The produced prediction models were able to outperform an uninformed guess but performed worse than models in prior studies. Achieving robust prediction models proved challenging for example due to misplaced data. Furthermore, due to underrepresentation of subjects within groups during the training of the models, the prediction models exhibited poor sensitivity. Multiple points of improvement were characterized during the study. Further research in the topic is required. A robust prediction model relying on PET image derived data rather than the images themselves will decrease the required resources in the study of memory problems. Over and above that, the features extracted will contribute to a better understanding of the pathophysiology of neurodegeneration.