Parameter Combination Optimization in Three-Dimensional Convolutional Neural Networks and Transfer Learning for Detecting Alzheimer's Disease from Magnetic Resonance Images

Alzheimer's disease (AD) destroys neurons in the brain, engendering brain atrophy and severely compromising brain function. Magnetic resonance imaging (MRI) is widely applied to analyze brain degeneration. AD is typically detected by examining specialist-selected features of two-dimensional images or region-of-interest features identified by trained classifiers. We developed a Taguchi-based three-dimensional convolutional neural network (T-3D-CNN) model for detecting AD in magnetic resonance images. CNN parameters are generally obtained through trial-and-error methods. To stabilize the CNN diagnostic accuracy, the Taguchi method was employed for parameter combination optimization. Obtaining patient data is difficult; thus, we performed transfer learning to verify the proposed T-3D-CNN model's effectiveness by using only a small quantity of patient data from various databases. The experimental results confirmed that the T-3D-CNN model detected AD from images in the Open Access Series of Imaging Studies (OASIS)-2 data set with an accuracy of 99.46%, which was 2.06 percentage points higher than that of the original 3D-CNN. After a complete investigation of the OASIS-2 data set, we selected 10, 30, 60, 80, and 100% of the data from the OASIS-1 data set to verify the T-3D-CNN and updated the original network weights through transfer learning; the average accuracies were 81.31, 92.88, 95.85, 100, and 100%, respectively.