Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Self-Supervised Pretext Tasks for Alzheimer's Disease Classification using 3D Convolutional Neural Networks on Large-Scale Synthetic Neuroimaging Dataset

Chen Zheng

TL;DR

Evaluated unsupervised methods to train a feature extractor for downstream AD vs. CN classification achieved similar performance compared to the same model using real-world data, which supports the feasibility of utilising large-scale synthetic data for pretext task training.

Abstract

Structural magnetic resonance imaging (MRI) studies have shown that Alzheimer's Disease (AD) induces both localised and widespread neural degenerative changes throughout the brain. However, the absence of segmentation that highlights brain degenerative changes presents unique challenges for training CNN-based classifiers in a supervised fashion. In this work, we evaluated several unsupervised methods to train a feature extractor for downstream AD vs. CN classification. Using the 3D T1-weighted MRI data of cognitive normal (CN) subjects from the synthetic neuroimaging LDM100K dataset, lightweight 3D CNN-based models are trained for brain age prediction, brain image rotation classification, brain image reconstruction and a multi-head task combining all three tasks into one. Feature extractors trained on the LDM100K synthetic dataset achieved similar performance compared to the same model using real-world data. This supports the feasibility of utilising large-scale synthetic data for pretext task training. All the training and testing splits are performed on the subject-level to prevent data leakage issues. Alongside the simple preprocessing steps, the random cropping data augmentation technique shows consistent improvement across all experiments.

Self-Supervised Pretext Tasks for Alzheimer's Disease Classification using 3D Convolutional Neural Networks on Large-Scale Synthetic Neuroimaging Dataset

TL;DR

Evaluated unsupervised methods to train a feature extractor for downstream AD vs. CN classification achieved similar performance compared to the same model using real-world data, which supports the feasibility of utilising large-scale synthetic data for pretext task training.

Abstract

Structural magnetic resonance imaging (MRI) studies have shown that Alzheimer's Disease (AD) induces both localised and widespread neural degenerative changes throughout the brain. However, the absence of segmentation that highlights brain degenerative changes presents unique challenges for training CNN-based classifiers in a supervised fashion. In this work, we evaluated several unsupervised methods to train a feature extractor for downstream AD vs. CN classification. Using the 3D T1-weighted MRI data of cognitive normal (CN) subjects from the synthetic neuroimaging LDM100K dataset, lightweight 3D CNN-based models are trained for brain age prediction, brain image rotation classification, brain image reconstruction and a multi-head task combining all three tasks into one. Feature extractors trained on the LDM100K synthetic dataset achieved similar performance compared to the same model using real-world data. This supports the feasibility of utilising large-scale synthetic data for pretext task training. All the training and testing splits are performed on the subject-level to prevent data leakage issues. Alongside the simple preprocessing steps, the random cropping data augmentation technique shows consistent improvement across all experiments.
Paper Structure (18 sections, 11 figures, 6 tables)

This paper contains 18 sections, 11 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Subjects
  5. Subject Selection
  6. Training Data
  7. Testing Data
  8. Preprocessing
  9. Pretext Tasks
  10. Brain Age Prediction
  11. Rotation Classification
  12. Image Reconstruction
  13. Multi-Head Task
  14. Representation Generation
  15. Experimental Settings
  16. ...and 3 more sections

Figures (11)

  • Figure 1: Comparison of brain tissue between a normal brain (left) and severe brain atrophy caused by Alzheimer's Disease (right) from a coronal (frontal) point of view. This figure is copied from website brainshrinkage2018.
  • Figure 2: MRI is defined by the plane (direction) of the image that is taken. Typically, three planes are used to describe the standard anatomical position of a human mri_planes. The basic orientation terms for a MRI of the body: from the inferior (I) to superior (S) is the axial plane; from the left (L) to right (R) is the sagittal plane; and from the anterior (A) to posterior (P) is the coronal plane. Figure source planes.
  • Figure 3: Three examples of T1-weighted MRI scans slices of subjects from the OASIS-3 oasis3 dataset in three planes: axial, coronal and sagittal. Row (a) shows images from a healthy subject whereas row (b) comes from a MCI subject with obvious brain changes in all planes. Lastly, row (c) reveals the severe brain tissue loss of an Alzheimer's subject.
  • Figure 4: Overview of the train-test-split.
  • Figure 5: Visualisation of each preprocessing step. Row (a) shows the raw image ($176\times256\times256$ voxels) of a 3D MRI scan whereas Row (b) presents the resized images ($192\times192\times192$ voxels). Row (c) shows the max-min intensity normalised image followed by row (d) of images that are processed by contrast limited adaptive histogram equalisation clahe.
  • ...and 6 more figures