Multi-class classification of Alzheimer’s disease through distinct neuroimaging computational approaches using Florbetapir PET scans

Abstract

Alzheimer’s disease (AD) is a neurological memory loss syndrome that eventually leads to incapacity to perform everyday chores and death. Since no known cure for this disease exists, it’s crucial to catch it early, before symptoms appear. This study used Florbetapir PET (specifically AV-45 PET scans) as a neuroimaging biomarker to develop a 3-Dimensional Ensemble Net for Alzheimer’s multi-class categorization. Our research is the first to examine the outcomes of three different feature extraction methods in depth (3D Subject, 3D Slice and 3D Patch Extraction Approach). Alzheimer detection through AV45 PET scans and 3-Dimensional slice neuroimaging computation technique was further done using three distinct Slicing algorithms (Subset Slice Algorithm (SSA), Uniform Slice Algorithm (USA), and Interpolation Zoom Algorithm (IZA). We tested the classification accuracy of a 3D patch-based technique with numerous patches varying from small to medium to huge dimensions. In this study, we used Amyloid- Positron Emission Tomography (AV45-PET) scans from the ADNI repository to create 3-Dimensional Ensemble Net model. Averaging, registering according to a standard template, and skull removal were used to pre-process the raw AV45-PET scans. The rotation method was used to augment these scans even more. Ensembling of two separate 3D-ConvNets was done for the 3D Subject-based computation technique. For the 3D Patch based computing approach, many non-overlapping patches ranging from 32, 40, 48, 56, 64, 72, 80, and 88 were retrieved and given to the Ensemble Net. Three unique algorithms were devised to extract slices from an AV45-PET scan and integrate them back to form a 3D volume in the 3D Slice based technique. Our results showed that (1) The three-class classification accuracy of our Ensemble Net model utilizing AV-45 PET images was 92.13% (maximum accuracy attained so far as per our knowledge). (2) The 3-Dimensional Patch extraction proposition was most accurate in Alzheimer’s categorization using Florbetapir PET images, followed by Subject-approach, then 3D Slice approach, with performance accuracy of 92.13, 91.01, and 90.44%, respectively. (3) The accuracy of the Ensemble Net network employing the 3D Patch computational approach was highest for larger patches (Dimensions as 72, 80, 88), next moderate patches (Dimensions as 56, 64, 48), and finally smaller patches (Dimensions as 32, 40). Higher dimension patches were classified correctly 92.13% of the time, whereas medium patches were correctly classified 80.89% of the time, and small patches were classified 74.63% of the time. (4) In terms of three-class classification accuracy, using 3D-Slice based approach, uniform slice extraction and interpolation zoom technique with 90.44 and 88.2% accuracy outperformed subset slice selection, with 81.46% performance accuracy. Using the Ensemble Net model and a 3D patch-based feature extraction approach, we efficiently labeled Alzheimer’s disease in a three-class categorization using amyloid PET scans with an accuracy of 92.13. Accuracy. The proposed model is examined using several neuroimaging feature computation approaches, stating that 3D-patch based Ensemble Net outperforms 3D-subject based Ensemble Net model and 3D-slice level model in terms of performance accuracy. In addition, a series of experiments were conducted for 3D-patch based approach with numerous patch dimensions varying from small to moderate to big sized patches in order to investigate impact of patch size on Alzheimer’s classification accuracy. While for 3D-Slice based approach, to determine which strategy was optimal, the slice-level technique was evaluated using three distinct algorithms revealing that uniform slice method and interpolation selection method outperforms subset slice method.