henhao
Sign in
BioinformaticsbioRxiv
Preprint

DeepAD: Alzheimer’s Disease Classification via Deep Convolutional Neural Networks using MRI and fMRI

Saman SarrafiD, Danielle D. DeSouzaiD, John A. E. AndersoniD, Ghassem Tofighi, Alzheimer’s Disease Neuroimaging Initiative

Abstract

1To extract patterns from neuroimaging data, various techniques, including statistical methods and machine learning algorithms, have been explored to ultimately aid in Alzheimers disease diagnosis of older adults in both clinical and research applications. However, identifying the distinctions between Alzheimers brain data and healthy brain data in older adults (age > 75) is challenging due to highly similar brain patterns and image intensities. Recently, cutting-edge deep learning technologies have been rapidly expanding into numerous fields, including medical image analysis. This work outlines state-of-the-art deep learning-based pipelines employed to distinguish Alzheimers magnetic resonance imaging (MRI) and functional MRI data from normal healthy control data for the same age group. Using these pipelines, which were executed on a GPU-based high performance computing platform, the data were strictly and carefully preprocessed. Next, scale and shift invariant low- to high-level features were obtained from a high volume of training images using convolutional neural network (CNN) architecture. In this study, functional MRI data were used for the first time in deep learning applications for the purposes of medical image analysis and Alzheimers disease prediction. These proposed and implemented pipelines, which demonstrate a significant improvement in classification output when compared to other studies, resulted in high and reproducible accuracy rates of 99.9% and 98.84% for the fMRI and MRI pipelines, respectively. Additionally, the subject-level classification was performed that resulted in the averaged accuracy rate of 94.32% and 97.88% for the fMRI and MRI pipeline respectively. Finally, a decision making algorithm was designed for the subject-level classification and improved the averaged accuracy rate to 97.77% for fMRI and 100% MRI subjects.

PDFDOI
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
332
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 17 Oct 2026 · from the full text of version 4 (14 Jan 2017) · CC BY-NC 4.0

Caffe LeNet and GoogleNet trained on 2D slices from ADNI scans of adults over 75 report 98.84% to near-perfect accuracy, and 100% per subject for MRI after majority voting.

Telling Alzheimer's brains from normally ageing ones over 75 is hard because atrophy patterns and image intensities overlap, and earlier deep-learning work paired autoencoders with support vector machines or a 3D CNN on structural MRI and PET, reporting accuracies such as 95.9%, 95.39% and 86.86%. Here whole preprocessed volumes are sliced into 2D images and fed to off-the-shelf CNN architectures, with resting-state fMRI used for this purpose for the first time according to the authors.

Key results

  1. ShownOn individual fMRI slices the two CNNs separate patients from controls almost perfectly.

    Across five random shuffles of the 4D time courses, the adopted LeNet averaged 99.9986% test accuracy on 28x28 inputs; GoogleNet, run on the first shuffle at 256x256, reached an accuracy of 100% as logged by Caffe. 793,800 PNG slices (270,900 Alzheimer's, 522,900 control) were split 75% training and 25% testing. Results were replicated with NVIDIA DIGITS Caffe on the same server. · 4.1 rs-fMRI Pipeline

  2. ShownStructural MRI slices classify slightly less well, and accuracy depends on how much spatial smoothing the preprocessing applied.

    Best LeNet accuracy 98.79% and best GoogleNet accuracy 98.8431%, both on the dataset smoothed with a Gaussian sigma of 3 mm; on unsmoothed data GoogleNet dropped to 84.5043%, below LeNet on the same images. Each MRI dataset held 62,335 slices (52,507 Alzheimer's, 9,828 control), and the LeNet runs were repeated five times. · 5.2 Structural MRI Pipeline; 4.2 Structural MRI Data Preprocessing

  3. ShownWhen whole subjects rather than slices are held out, accuracy falls but a majority vote over each subject's slices restores it.

    Shuffling subjects five times and assigning 75% to training gave average subject-level accuracy of 94.32% for fMRI and 97.88% for MRI; a majority-rule vote over the slices of each subject raised these to 97.77% and 100%. Subject-level results were best for MRI smoothed at sigma 3 mm and 4 mm, which the authors contrast with the slice-level experiment. · 6 Subject-level recognition; 7 Decision making: Majority rule; Abstract

  4. ShownThe five-to-one class imbalance in the MRI data does not explain the high accuracy.

    Retraining LeNet after cutting Alzheimer's images to 10,722 against 9,828 control images lowered the best accuracy by about 1% (97.81% for the balanced sigma 3 mm dataset), so going from roughly 5:1 to 1:1 left classification essentially unchanged. · 5.2 Structural MRI Pipeline

Methods, data and reproducibility

Resting-state fMRI and structural MRI from ADNI were preprocessed with FSL (brain extraction, motion and slice-timing correction, registration to standard space; grey-matter templates and smoothing at sigma 2, 3 and 4 mm for MRI), decomposed into 2D slices saved as PNG, and used to train LeNet and GoogleNet in Caffe for binary classification, with a majority vote across slices for subject-level calls.

  • ADNI resting-state fMRI subset, age group over 75 · human · n = 144 subjects: 52 Alzheimer's, 92 healthy controls
  • ADNI structural MRI subset, age group over 75 · human · n = 302 subjects: 211 Alzheimer's, 91 healthy controls
  • PNG slices generated from the fMRI time courses for training and testing · human · n = 793,800 images (270,900 Alzheimer's, 522,900 control)
  • PNG slices per structural MRI dataset (unsmoothed plus three smoothing kernels) · human · n = 62,335 images each (52,507 Alzheimer's, 9,828 control)
Methods & tools
Caffe, NVIDIA DIGITS, LeNet, GoogleNet, FSL (BET, MCFLIRT, VBM), dcm2nii, Nibabel, OpenCV, LMDB
Reagents & environment
Caffe with LMDB inputs, repeated under NVIDIA DIGITS Caffe; 30 epochs of stochastic gradient descent (gamma 0.1, momentum 0.9, base rate 0.01, step policy), weight decay 0.005 for the fMRI LeNet and 0.0005 elsewhere; inputs 28x28 for LeNet, 256x256 for GoogleNet. No package versions given.
Assay conditions & compute
Amazon AWS Linux G2.8xlarge instances with four NVIDIA GPUs (1,536 CUDA cores and 4GB video memory each) and 32 Intel Xeon E5-2670 vCPUs with 60 GB memory; no training times are reported.
Reproducibility gaps
No code repository or model weights are mentioned. The imaging comes from the ADNI database, disseminated by the Laboratory for Neuro Imaging at USC, with no URL or accession in this text; the link for the FSL-VBM protocol is also missing.
Limitations (4)
  • At slice level the images were shuffled before the 75/25 split, so slices from the same subject and time course appear in both training and testing; the drop to 94.32% and 97.88% when subjects are held out indicates how much that inflates slice-level figures. · noted in this summary
  • Everything comes from one ADNI cohort over 75 with no external or multi-site held-out validation, and some subjects contributed scans from several time points treated separately. · noted in this summary
  • Only Alzheimer's versus healthy control is attempted; predicting stages of Alzheimer's progression is left to future studies. · stated by the authors
  • Subject-level accuracy drops because training sees less variety, and the decision rule has no tie-breaker beyond deferring to clinical tests. · stated by the authors

Who should read:Read this if you build imaging classifiers for neurodegenerative disease and want a concrete slice-based CNN pipeline, or a cautionary example of how slice-level and subject-level accuracy diverge.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISSarraf '16Qiu '20Feng '18Gupta '19Lu '22Dhinagar '23
90% semantic matchBioinformatics
Saman Sarraf, Ghassem Tofighi·2016 Future Technologies Conference (FTC)·Dec 1, 2016