P1‐119: ENHANCING DEEP LEARNING MODEL PERFORMANCE FOR AD DIAGNOSIS USING ROI‐BASED SELECTION

• DOI: 10.1016/j.jalz.2019.06.674
• Publication Date: 2019-10-18T10:44:06Z
• Authors: Shangran Qiu; Megan S. Heydari; Matthew I. Miller; Prajakta S. Joshi; Benjamin C. Wong; Rhoda Au; Vijaya B. Kolacha...
• Keywords: Region of interest; Binary classification

Utilization of 3D neuroimaging data for AD diagnosis has been great interest in recent years. Deep learning algorithms based on convolutional neural networks (CNN) are increasingly being considered analyzing MR scans to diagnose cognitively impaired individuals. Most the CNN models focus using two-dimensional (2D) slices or whole volume-based three-dimensional (3D) as inputs build binary classification models. We leveraged a series 2D and developed region (ROI)-based strategy enhance accuracy model (normal vs AD). Volumetric MRI 416 subjects from Alzheimer's Disease Neuroimaging Initiative (ADNI-1) were linear registered template. A developed, each built upon different slice axial, coronal sagittal viewing planes classify cases normal controls. Within brain, we then identified prismatic volume containing plane that yielded above threshold accuracy, constructed this input data. The entire ADNI dataset was randomly divided into 3:1:1 ratio, where 60% used training, 20% validation remaining testing. Models evaluated sensitivity specificity. Performance varied function spatial location (Figure 1; Table 1). derived ROI, comprised accuracies 70% qualitatively associated with brain regions linked attention, memory executive 2). selected ROI resulted an enhanced 90.0%, surpassing individual performances (Table 1),