研究目的
To analyze amyloid-PET images with florbetapir tracer for detecting Alzheimer's disease or Mild Cognitive Impairment stages using deep learning methods.
研究成果
The developed pipeline and deep neural network achieved a sensitivity of 90% and specificity of 78% for classifying AD stages using features from 18F-AV-45 and 18F-FDG PET images. This provides adequate information for physicians in disease diagnostics, with future work focused on improving model parameters and topology to increase accuracy.
研究不足
The main limitation is the half-time period of the radioactive tracer, though florbetapir's longer half-time (110 minutes) mitigates this compared to other tracers like 11C-PiB. There is significant overlapping between classes in the data, and the model's accuracy could be improved with more images or additional biologically meaningful features.
1：Experimental Design and Method Selection:
The study uses a pipeline for processing PET images, including spatial normalization and feature extraction, followed by developing a deep neural network for multiclass classification of AD stages (HC, MCI, AD).
2：Sample Selection and Data Sources:
Image data (18F-AV-45, 18F-FDG PET, and structural MRI) were acquired from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database, with baseline information for 731 participants labeled as Normal, MCI, or AD based on MMSE and CDR scores.
3：List of Experimental Equipment and Materials:
PET scanners used include General Electric (GE) Discovery, Philips Gemini, Siemens HRRT 207-slice, and BioGraph HiRez (Model 1093). Software tools include SPM Toolbox for image processing and AAL atlas for brain region mapping.
4：3). Software tools include SPM Toolbox for image processing and AAL atlas for brain region mapping.
Experimental Procedures and Operational Workflow:
4. Experimental Procedures and Operational Workflow: Images were processed by updating origins, spatial alignment, normalization (using cerebellum as a reference region for SUVr calculation), and feature extraction (mean voxel intensity from 116 brain regions). Features were reduced using PCA, and a neural network with six hidden layers (using ReLU and Softmax activation functions) was trained with minibatches and L2 regularization.
5：Data Analysis Methods:
The neural network was trained, tested, and validated with data splits (90% train, 5% test, 5% validation), and performance was evaluated using sensitivity and specificity metrics.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
