| Summary: | 碩士 === 國立交通大學 === 統計學研究所 === 107 === Parkinson’s disease (PD) is a long-term degenerative disorder of central nervous system that is prevalent in elderly. It is currently medically diagnosed by functional medical imaging. The typical types of functional brain imaging include Positron emission tomography (PET) and Single Photon Emission Computed Tomography (SPECT).
In this study, we use a dataset containing 202 SPECT imaging which is consisted of 6 normal healthy controls and 196 patients with PD. In addition, according to the severity of illness, PD can be divided into 5 stages. The statistical models are used for quantitative analysis, and provide more references for the clinical diagnosis of PD.
Used statistical models include traditional models, ensemble models, and deep learning models. The goal is to make a good prediction of the PD illness stages. First, we select the slice whose striatum can be recognized most. For traditional and ensemble models, there are three kinds of feature extraction method to be used, including PCA, MPCA, and image statistics. Furthermore, we use the Laws' Texture Energy Measure (LTEM) method to do a further analysis of the imaging, which will extend the numbers of features. The best combination of parameters is found by grid search. We use cross validation to evaluate the model performance. For deep learning models, we use the technique of image augmentation to increase the data size, and build model by the architecture of VGG16. Also, we use the Auto-ML to build a model, which is a state-of-the-art model that can generate a neural architecture automatically.
The results show that, in traditional and ensemble models, the image statistics approach is the best feature extractor of the three, and the random forest model outperforms other approaches. Additionally, LTEM method might be helpful to get the features of an image. Overall, the deep learning VGG16 model has the best performance without any further image preprocessing. It is found that the VGG16 model can capture significant features from imaging, reaching a higher classification accuracy.
|
|---|
