Source-linked AI summary
Classification of Alzheimer's Disease using fMRI Data and Deep Learning Convolutional Neural Networks
Saman Sarraf, Ghassem Tofighi
TL;DR
The paper addresses the challenge of selecting discriminative features for classifying Alzheimer’s disease from healthy controls using fMRI data. It applies a CNN based on the LeNet architecture to labeled images, achieving 96.86% accuracy, while noting computational complexity and infrastructure costs.
Problem
Classifying Alzheimer’s clinical data is challenging because selecting the most discriminative features is difficult, yet distinguishing disease from healthy subjects could support disease recognition.
Method
The study converts fMRI data into labeled 2D images and uses a CNN-based LeNet model for binary Alzheimer’s-versus-Normal classification.
Results
96.86% accuracy was achieved for Alzheimer’s data classification, compared with 84% reported for SVM in the literature.
Takeaways & Limitations
The CNN-based pipeline combines feature extraction and classification and could potentially be extended to predict different Alzheimer’s disease stages and age groups.
Abstract
from arXiv · showhide
Over the past decade, machine learning techniques especially predictive modeling and pattern recognition in biomedical sciences from drug delivery system to medical imaging has become one of the important methods which are assisting researchers to have deeper understanding of entire issue and to solve complex medical problems. Deep learning is power learning machine learning algorithm in classification while extracting high-level features. In this paper, we used convolutional neural network to classify Alzheimer's brain from normal healthy brain. The importance of classifying this kind of medical data is to potentially develop a predict model or system in order to recognize the type disease from normal subjects or to estimate the stage of the disease. Classification of clinical data such as Alzheimer's disease has been always challenging and most problematic part has been always selecting the most discriminative features. Using Convolutional Neural Network (CNN) and the famous architecture LeNet-5, we successfully classified functional MRI data of Alzheimer's subjects from normal controls where the accuracy of test data on trained data reached 96.85%. This experiment suggests us the shift and scale invariant features extracted by CNN followed by deep learning classification is most powerful method to distinguish clinical data from healthy data in fMRI. This approach also enables us to expand our methodology to predict more complicated systems.
I. INTRODUCTION
Alzheimer’s disease progressively damages brain cells and cognitive abilities, while rs-fMRI offers a non-invasive way to measure functional brain changes. The paper addresses the clinical need for algorithms that distinguish Alzheimer’s data from healthy subjects.
- Alzheimer’s disease is an irreversible, progressive disorder that destroys brain cells and impairs memory, thinking, and daily functioning.
- Resting-state fMRI measures functional brain activity and brain changes non-invasively without requiring patients to perform a task.
- Classifying fMRI data could help clinicians and scientists distinguish Alzheimer’s disease from healthy subjects.
A. Data Acquisition and Preprocessing
The study selected Alzheimer’s patients and healthy older controls from the ADNI dataset and acquired anatomical and functional MRI scans using a 3 Tesla Siemens Trio scanner.
- Participants included Alzheimer’s patients and 15 elderly normal controls selected from the ADNI dataset.The controls included 24 females and 19 males with a mean age of 74.9 ± 5.7 years.
- Normal participants were healthy without reported medical or neurological conditions, while Alzheimer’s patients had MMSE scores above 20.
- Scanning used a Siemens Trio 3 Tesla MRI system with 3D MP-RAGE anatomical and EPI functional sequences.
B. Deep Learning
Deep learning is presented as a hierarchical machine-learning approach inspired by the brain that extracts high-level abstractions through complex neural-network architectures.
- Hierarchical deep learning models high-level features and abstractions from data using complex neural-network architectures.
- The approach was motivated partly by the neocortex’s hierarchical processing of sensory signals over time.
- Deep learning is described as a modern branch of machine learning inspired by the human brain.
1) Convolutional Neural Networks (CNNs / ConvNets):
CNNs process two-dimensional data through spatially structured layers that learn local features and support classification. Their architecture can provide shift, scale, and rotation invariance while reducing learned parameters.
- Convolutional Neural Networks (CNNs / ConvNets): CNNs assume two-dimensional raw inputs, exploit spatial relationships, and reduce the number of parameters requiring learning.
- Convolutional Neural Networks (CNNs / ConvNets): Local receptive fields let CNN neurons detect elementary features such as oriented edges or corners.
- Convolutional Neural Networks (CNNs / ConvNets): CNN architectures provide invariance to shifts, scale, and rotation through hierarchical processing of local features.
- Convolutional Neural Networks (CNNs / ConvNets): Convolutional layers learn spatially positioned features using small filters that share parameters across activation maps.
- Convolutional Neural Networks (CNNs / ConvNets): Fully connected layers compute class scores from the preceding volume.
2) Adoped LeNet-5:
The study adopts LeNet-5, originally designed for handwritten-digit classification, for the binary classification of Alzheimer’s and normal fMRI data.
- LeNet-5 was originally designed for handwritten-digit classification and check-number recognition.
- The architecture was adapted to classify Alzheimer’s and normal data as a binary problem.
- The implemented LeNet-5 network is presented in the paper’s architecture figures.
III. RESULTS AND DISCUSSION
The study converts preprocessed fMRI data into labeled 2D images for LeNet-based classification, evaluates performance through repeated validation, and reports 96.8588% averaged accuracy. The discussion highlights CNN’s combined feature-extraction and classification capabilities while noting computational costs.
- Data preparation: Preprocessed 4D Nifti fMRI data were converted into labeled 2D JPEG images and stored in LMDB databases for deep-learning classification.
- Data preparation: The data were divided into training (60%), validation (20%), and testing (20%) subsets.
- Evaluation: Five-fold cross-validation was repeated five times to support robustness and reproducibility, with results gathered in Table I.
- Results: 96.8588% averaged accuracy was achieved in recognizing Alzheimer’s data from normal controls.
- Training: The learning rate started at 0.01 and was divided by 10 every 10 epochs.
- Discussion: CNN combines high-level feature extraction with classification and improved reported accuracy from 84% using SVM to 96.86%.The paper also identifies high algorithmic complexity and expensive infrastructure as problems of deep-learning solutions.
IV. CONCLUSIONS
The study classified Alzheimer’s disease data from normal controls with high accuracy using a CNN-based LeNet architecture. The authors propose extending the approach with more convolutional layers for more complicated problems.
- IV. CONCLUSIONS: 96.86% accuracy was achieved when CNN deep learning architecture using LeNet classified Alzheimer’s disease data from normal controls.The network was trained and tested with a large number of images.
- IV. CONCLUSIONS: The proposed solution potentially supports prediction of new data and generalization to different Alzheimer’s disease stages and age groups.The paper also describes the architecture as enabling feature selection and classification.
- IV. CONCLUSIONS: The first and second convolution layers are presented with statistics and visualizations for Alzheimer’s disease samples.
- IV. CONCLUSIONS: More complicated network architectures with additional convolutional neural layers are recommended for future work and more complicated problems.
VI. APPENDIX I
The appendix acknowledges that the project’s data collection and sharing were funded through ADNI-related public, departmental, institutional, and private-sector contributions.
- VI. APPENDIX I: Data collection and sharing for the project were funded by the Alzheimer’s Disease Neuroimaging Initiative.
- VI. APPENDIX I: Funding included NIH grant U01 AG024904 and Department of Defense ADNI award W81XWH-12-2-0012.
- VI. APPENDIX I: ADNI funding came from the National Institute on Aging, the National Institute of Biomedical Imaging and Bioengineering, and listed private-sector contributors.