College - Author 1
College of Engineering
Department - Author 1
Computer Engineering Department
Degree Name - Author 1
BS in Computer Engineering
Date
6-2022
Primary Advisor
Xiao-Hua Yu, College of Engineering, Electrical Engineering Department
Abstract/Summary
The objective of this project was to evaluate multiple classification methods for Alzheimer’s Detection. The approach developed includes preprocessing, feature extraction, and classification. Preprocessing included normalization and skull extraction. The wavelet transform was used for feature extraction. Three wavelet types were evaluated at three levels of decomposition. The resulting coefficients from the wavelet transform were included as the input to a simple Convolutional Neural Network (CNN) model for classification. From the multiple, CNN models assessed, the performance results of three different classification models were noted in this document. Two datasets were used for training and classification with the proposed methodology. The first dataset was from Kaggle and represented a subset of the Open Access Series of Imaging Studies (OASIS), and the second dataset was from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). The Kaggle dataset had four classifications, and the ADNI dataset had three classifications. All CNN models were modified for the different dataset to match the specific classification labels. The classification method provided in the paper was developed in Python. The skull extraction was performed using OpenCV’s connected components computation and closing morphology transformation. The wavelet transform was computed using the PyWavelets library and all CNN models were implemented using the Keras deep learning library. The approach with the optimal performance utilized the Haar wavelet transform with two levels of decomposition. The optimal CNN model included two convolutional layers, a dropout rate of 0.1, and a batch size of 10. The optimal model was determined by four performance metrics. Those metrics include accuracy, precision, recall, and specificity. The accuracy of the optimal model was 0.791, the precision was 0.664, the recall was 0.335, and the specificity was 0.994.
URL: https://digitalcommons.calpoly.edu/cpesp/337