[Paper Review] Ensemble learning with 3D convolutional neural networks for connectome-based prediction

Meenakshi Khosla, Keith Jamison+2 more|arXiv (Cornell University)|Sep 11, 2018

Functional Brain Connectivity Studies51 references3 citations

TL;DR

This paper proposes an ensemble learning framework using 3D convolutional neural networks (3D-CNNs) to improve functional connectome-based prediction by combining models trained on diverse brain parcellations. It demonstrates that stochastic parcellations yield performance comparable to standard atlases, and the ensemble strategy significantly enhances robustness and accuracy in classifying autism spectrum disorder and predicting age from rs-fMRI data.

ABSTRACT

The specificty and sensitivity of resting state functional MRI (rs-fMRI) measurements depend on pre-processing choices, such as the parcellation scheme used to define regions of interest (ROIs). In this study, we critically evaluate the effect of brain parcellations on machine learning models applied to rs-fMRI data. Our experiments reveal a remarkable trend: On average, models with stochastic parcellations consistently perform as well as models with widely used atlases at the same spatial scale. We thus propose an ensemble learning strategy to combine the predictions from models trained on connectivity data extracted using different (e.g., stochastic) parcellations. We further present an implementation of our ensemble learning strategy with a novel 3D Convolutional Neural Network (CNN) approach. The proposed CNN approach takes advantage of the full-resolution 3D spatial structure of rs-fMRI data and fits non-linear predictive models. Our ensemble CNN framework overcomes the limitations of traditional machine learning models for connectomes that often rely on region-based summary statistics and/or linear models. We showcase our approach on a classification (autism patients versus healthy controls) and a regression problem (prediction of subject's age), and report promising results.

Motivation & Objective

• To evaluate the impact of different brain parcellation schemes on machine learning performance in functional connectome-based prediction.
• To address the sensitivity of connectome models to arbitrary ROI definitions by proposing an ensemble learning strategy across multiple parcellations.
• To develop a 3D-CNN framework that leverages full-resolution 3D spatial structure of rs-fMRI data to capture non-linear patterns.
• To improve prediction accuracy for clinical applications such as autism classification and age regression using connectome data.
• To demonstrate that stochastic parcellations can match or exceed performance of standard atlases, enabling robust ensemble learning.

Proposed method

• The method employs stochastic parcellations via Poisson Disk Sampling to generate diverse region-of-interest (ROI) definitions across subjects.
• For each stochastic parcellation, functional connectivity matrices are computed using Pearson correlation between ROIs.
• A 3D-CNN model is trained on the 3D volumetric representation of rs-fMRI data, preserving spatial resolution and capturing local patterns.
• An ensemble strategy combines predictions from multiple 3D-CNN models trained on different parcellations, improving generalization.
• The framework is evaluated using ABIDE-II data with 10,000 bootstrap samples to assess stability and performance across models.
• The approach avoids reliance on hand-crafted features or linear models, instead learning hierarchical representations directly from 3D fMRI volumes.

Experimental results

Research questions

• RQ1 How do different brain parcellation schemes affect the performance of machine learning models in functional connectome-based prediction?
• RQ2 Can stochastic parcellations yield prediction performance comparable to established atlases in connectome-based classification and regression tasks?
• RQ3 Does an ensemble of models trained on diverse parcellations improve robustness and accuracy compared to single-atlas models?
• RQ4 To what extent does a 3D-CNN architecture that preserves full 3D spatial structure outperform traditional linear or graph-based models on connectome data?
• RQ5 How does the proposed ensemble strategy compare to baseline models (e.g., SVM, FCN, BrainNet-CNN) in terms of predictive performance and stability?

Key findings

• Stochastic parcellations consistently achieved performance comparable to widely used atlases at the same spatial scale, indicating their reliability for model training.
• The SP-Ensemble approach achieved classification accuracy on the autism vs. healthy controls task that matched or slightly exceeded the best single-atlas model.
• The 3D-CNN model outperformed baseline models (e.g., FCN, SVM) in age regression, with more pronounced improvements for individual atlas models.
• Ensemble predictions showed marginal differences between models, suggesting that the ensemble strategy stabilizes performance across different parcellation schemes.
• Kernel density estimates from 10,000 bootstrap samples showed that the proposed 3D-CNN and SP-Ensemble approaches underperformed competing methods in less than 50% of samples, indicating strong consistency.
• The study found that the choice of correlation metric (e.g., Pearson) remains an arbitrary factor, though it did not significantly alter the core findings.