Siamese Network for Content-Based Image Retrieval: Detection of Alzheimer's Disease from neuroimaging data

Abstract

In recent years deep-learning methods have demon-strated impressive results in various domains of computer vision, including medical imaging. This paper examines the possibility of leveraging deep-learning concepts in designing a computer system that could help clinicians make accurate Alzheimer disease (AD) diagnosis by retrieving the most similar archived brain scans of patients with already known diagnoses. We implement a siamese network with ResNet-50 twin subnetworks and train it on the MRI data obtained from ADNI (Alzheimer's Disease Neu-roimaging Initiative) dataset. Four different approaches for slice extraction from MRI volume are considered: using the three slices from the same plane (axial, coronal or sagittal) and combining one slice from each plane. The final performance of the CBIR system on new patient's data based only on MR neuroimaging modality shows limited and comparable performance with all four approaches and leaves space for further enhancements, including complementing neuroimaging MRI data with other data modalities relevant for AD detection.