Multiplane Convolutional Neural Network (Mp-CNN) for Alzheimer’s Disease Classification

One of the objectives of medical imaging research is to develop an effective and reliable clinical support tool for the early detection of various neurological conditions in patients such as Alzheimer's Disease (AD).Classification based on three-dimensional (3D) MRI (Magnetic-Resonance-Imaging) images have shown a goodperformance. However, solving 3D object classification as a 3D object using classical machine learning or deeplearning is computationally high. 3D images from MRI can be reconstructed into three depth directions: axial, coronal,and sagittal. However, the multi-view-based method cannot explore the content of each image slice for all planes in3D-MRI. Our study proposes a multiplane method by using three two-dimensional (2D) CNNs to capturediscriminatory information on all available planes. Our method is a new fusion strategy to deal with the disadvantagesof shape-based multi-view techniques. Our framework selects three slices: the three largest 3-planes to represent whole3D-MRI objects as multi-2DCNN inputs. Extensive experiments have shown that the proposed method can outperformthe 2DCNN method, which uses only one plane. More importantly, by taking full advantage of 2DCNN, we offer anew method for identifying 3D objects that is both easy and efficient. We called this new architecture MultiplaneConvolutional Neural Network (Mp-CNN) since it used multiple inputs in its design. We evaluated the proposedmethod using T1-weighted structural MRI data consisting of 500 AD, 500 Mild Cognitive Impairment (MCI), and 500Normal Cognition (NC) subjects collected from the MRI database of Alzheimer's Disease Neuroimaging Initiative(ADNI). From the performed experiments, the proposed method achieves 93% accuracy for multiclass AD-MCI-NCand good precision AD 93%, MCI 91%, and NC 95%, respectively.