BACKGROUND
Image-based recognition has become a long-term topic in the field of artificial intelligence, and neuroimaging has gradually become a beneficial way to understand the course of Alzheimer’s disease (AD).
OBJECTIVE
The goal of this study is to compare the detection performance of convolutional neural networks (CNNs) on medical images to establish a classification model for epidemiological research. However, medical image analysis lacks large labeled training datasets, and thus many transfer learning-based methods have been proposed to solve few labels in the medical field.
METHODS
Owing to the scarcity of image data from single-photon emission computed tomography (SPECT), this study uses transfer learning to compare the performance of diagnostic methods based on five different CNNs (two lightweight and three deeper-weight CNN models) to determine the most suitable model. Brain scan image data were collected from 36 male and 63 female subjects. This study used 4711 images as the input data for the model.
RESULTS
The Cognitive Abilities Screening Instrument and Mini Mental State Exam scores of subjects with Clinical Dementia Rating (CDR) of 2 were significantly lower than those of subjects with CDR of 1 and 0.5. These results indicate that the ResNet model (the deeper-weight CNN model) exhibits the highest accuracy (70.79%) and can hence be used to improve the classification of mild cognitive impairment (MCI), mild AD, and moderate AD (CDRs of 0.5, 1, and 2, respectively).
CONCLUSIONS
This study successfully analyzes the classification performance of different CNN architectures for medical images and also proves the effectiveness of transfer learning in identifying the MCI, mild AD, and moderate AD scoring based on SPECT images.