Problem motivation: Alzheimer's disease (AD) is the most common cause of dementia and estimated by World Health Organisation as the fifth most common cause of death in the world. The increasing proportion of elderly population, which is particularly pronounced in the developed countries, is predicted to double, compared to 2020 figures, the number of Alzheimer's dementia patients by 2050.
Currently treatment for AD is not available, but research shows that early diagnosis of this disease would importantly expedite the development of drugs and symptomatic therapies. Early diagnosis of AD is a difficult and therefore still open and important research question due to the relatively late development of AD signs. Cerebrospinal fluid (CSF) biomarkers and some imaging biomarkers such as positron emission tomography (PET) and magnetic resonance imaging (MRI) show great potential.
In the master's thesis we evaluated regression models based on the analysis of blood test data, PET and MRI images, with the aim to predict concentrations of CSF-extracted beta amyloid, tau protein (total-tau) and phospho-tau (pTau-181). In practice, these CSF values are usually determined by laboratory tests on CSF samples obtained by lumbar puncture (LP). Due to the invasiveness of LP and associated hazards, predicting the value of CSF parameters from a non-invasive test data - such as MRI imaging - is a clinically interesting idea and the central problem of investigation in this master's thesis.
Methods: regression of CSF parameters from 3D PET and MRI medical images from a total of seven datasets was performed with deep neural networks and, as the baseline, classical machine learning models were implemented using blood test results and Standard Uptake Value Ratio calculated from Pittsburg compound B PET images.
Results: as expected, regression models trained on the results of routine blood tests showed poor performance, which indicates that classical blood tests do not contain relevant information for predicting the values of the CSF parameters. More accurate CSF parameter regression models were based on PET images and SUVR ratios with an average relative absolute error of the order of 15 % of the range of the target CSF variable. Models performed worse at interval ends of the target variable. The cause is most likely a markedly uneven distribution of input data, making it difficult to train models of deep neural networks. Models trained on MRI images showed a slightly higher average regression error than PET images and with larger errors at target variable interval ends.
Conclusion: the developed regression models do not seem clinically applicable due to large deviations at target variable interval ends. In the intervals, where there were sufficient input and associated CSF data the regression error was about 30 percent with respect to the reference value, which could be clinically feasible given the dichotomous treatment of CSF value in the AD diagnosis domain. We feel that the developed regression models show potential for improved accuracy of prediction of CSF values in case additional data samples are provided from the underrepresented intervals of the target CSF variable.
|