Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Multi-Stage Graph Learning for fMRI Analysis to Diagnose Neuro-Developmental Disorders

Wenjing Gao, Yuanyuan Yang, Jianrui Wei, Xuntao Yin, Xinhan Di

TL;DR

A multi-stage graph learning framework which incorporates 1) pretrain stage : self-supervised graph learning on insufficient supervision of the fmri data 2) fine-tune stage : supervised graph learning for brain disorder diagnosis.

Abstract

The insufficient supervision limit the performance of the deep supervised models for brain disease diagnosis. It is important to develop a learning framework that can capture more information in limited data and insufficient supervision. To address these issues at some extend, we propose a multi-stage graph learning framework which incorporates 1) pretrain stage : self-supervised graph learning on insufficient supervision of the fmri data 2) fine-tune stage : supervised graph learning for brain disorder diagnosis. Experiment results on three datasets, Autism Brain Imaging Data Exchange ABIDE I, ABIDE II and ADHD with AAL1,demonstrating the superiority and generalizability of the proposed framework compared to the state of art of models.(ranging from 0.7330 to 0.9321,0.7209 to 0.9021,0.6338 to 0.6699)

Multi-Stage Graph Learning for fMRI Analysis to Diagnose Neuro-Developmental Disorders

TL;DR

A multi-stage graph learning framework which incorporates 1) pretrain stage : self-supervised graph learning on insufficient supervision of the fmri data 2) fine-tune stage : supervised graph learning for brain disorder diagnosis.

Abstract

The insufficient supervision limit the performance of the deep supervised models for brain disease diagnosis. It is important to develop a learning framework that can capture more information in limited data and insufficient supervision. To address these issues at some extend, we propose a multi-stage graph learning framework which incorporates 1) pretrain stage : self-supervised graph learning on insufficient supervision of the fmri data 2) fine-tune stage : supervised graph learning for brain disorder diagnosis. Experiment results on three datasets, Autism Brain Imaging Data Exchange ABIDE I, ABIDE II and ADHD with AAL1,demonstrating the superiority and generalizability of the proposed framework compared to the state of art of models.(ranging from 0.7330 to 0.9321,0.7209 to 0.9021,0.6338 to 0.6699)
Paper Structure (13 sections, 5 equations, 4 figures, 2 tables)

This paper contains 13 sections, 5 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Graph Network for fMRI
  4. Multi-Stage Graph Learning
  5. Methods
  6. Pre-training Stage
  7. Fine-tuning stage
  8. Experimental Evaluation
  9. Experimental setup
  10. Evaluation
  11. Ablations Study for ABIDE I , ABIDE II and ADHD.
  12. Model Interpretability
  13. Discussion

Figures (4)

  • Figure 1: Illustration of the proposed Multi-Stage Graph. The model can be divided into a hypergraph autoencoder (pre-training) and a hypergraph classification model (fine-tuning). (1) We propose a hypergraph autoencoder which consists of three stages: edge embedding learning and graph embedding learning. Multi-Stage Graph performs contrastive learning on features extracted from the original hypergraph and its edge-dropped version based on a Bernoulli mask, enabling the learned features to be independent of the class labels and genuinely represent an embedding of the graphs in Euclidean space. (2) The pre-trained hypergraph autoencoder and fMRI BOLD signals are then combined to obtain the final graph embedding and diagnose results under the supervision of the class labels.
  • Figure 2: Model interpretability on ABIDE I. (a) All 66 selected functional connection features shown in the circus plot. (b) 66 functional connections on the cortical surface. The red and blue lines in (a) and (b) represent the increased and decreased strength functional connections in ASD compared to TD, respectively. (c) Increased node features in ASD compared to TD. (d) Decreased node features in ASD compared to TD. In (c) and (d), the red, green, and blue bubbles represent the Slow-5, Slow-4, and classical ALFF features, respectively.
  • Figure 3: Model interpretability on ABIDE II. (a) All 66 selected functional connection features shown in the circus plot. (b) 66 functional connections on the cortical surface. The red and blue lines in (a) and (b) represent the increased and decreased strength functional connections in ASD compared to TD, respectively. (c) Increased node features in ASD compared to TD. (d) Decreased node features in ASD compared to TD. In (c), the red, green, and blue bubbles represent the Slow-5, Slow-4, and classical ALFF features, respectively.
  • Figure 4: Qualitative results of heatmaps generated by different baselines (the red mask represents lesion area, dark blue arrows represent various noise).