DeepPET-3D: A Deep Learning Based 3D-CNN Model for Diagnosis of Alzheimer's Disease Using 18-FDG-PET

Abstract

Diagnosis of Alzheimer’s disease is a challenging task, and detection can potentially help prevent its progression toward Dementia. Computer-aided techniques incorporating artificial intelligence have proven effective for diagnosing medical images. Deep learning tools like convolutional neural networks assist in extracting relevant visual information and thus avoiding manual feature extraction and interpretation. In this research, we propose a novel 3D-CNN model for the early detection of AD using positron emission tomography (PET). The dataset acquired from ADNI consists of 3D images of the brain segregated into three classes, namely Alzheimer’s Disease (AD), Mild-Cognitive Impairment (MCI), and Cognitive Normal (CN). After data acquisition, we performed various pre-processing methods like thresholding, normalization, volume-reduction, and image augmentation. This study performed two types of experiments: multi-class and binary classification. The multi-class classification achieved an accuracy of 92.31%. Furthermore, the accuracy for AD vs CN, CN vs MCI, and MCI vs AD was 94.79%, 93.28%, and 96.91%, respectively.

Dishant Padalia

Computer Science Master’s Student at UMass Amherst

My research interests include computer vision, AI in healthcare and natural language processing.