Supervised Diagnosis Prediction from Cortical Sulci: Toward the Discovery of Neurodevelopmental Biomarkers in Mental Disorders

Pierre Auriau, Antoine Grigis, Benoit Dufumier, Robin Louiset, Joël Chavas, Pietro Gori, Jean-François Mangin, Edouard Duchesnay

TL;DR

This paper explores using the intricate shapes of cortical sulci, formed early in brain development, as biomarkers for predicting mental disorders like autism, bipolar disorder, and schizophrenia. By applying deep learning models to structural MRI data, the researchers demonstrate that these early developmental features hold significant diagnostic information, paving the way for more personalized psychiatric approaches.

What Problem Does This Paper Address?

Mental disorders are complex, arising from a mix of genetic and environmental factors. Disentangling the neurodevelopmental origins of these conditions from later environmental influences is crucial for understanding their underlying mechanisms and developing targeted interventions. Traditional methods using voxel-based morphometry (VBM) on structural MRI (sMRI) often struggle to isolate these neurodevelopmental contributions, as they retain information from both sources. This paper aims to address this by focusing on cortical sulci, whose shapes are primarily shaped during early brain development, making them a promising proxy for neurodevelopmental abnormalities.

Methods

The core of this research involves training deep learning models, specifically Convolutional Neural Networks (CNNs), to predict psychiatric diagnoses based on the 3D shapes of cortical sulci.

  • Sulcus Skeletonization: sMRI T1w images are processed using the Brainvisa/Morphologist pipeline to extract "sulcus skeletons." These are essentially binary representations of the brain's folds, capturing their intricate geometry.
  • Image Pre-processing: To address the sparsity of skeleton images, Gaussian smoothing was explored. However, the study found that this did not significantly improve model performance and decided against its use in subsequent experiments.
  • CNN Architectures: Three established CNN architectures—AlexNet, ResNet18, and DenseNet121—were benchmarked for their diagnostic prediction capabilities. DenseNet showed the most consistent performance.
  • Training Strategies: Both supervised learning with a binary cross-entropy (BCE) loss and weakly supervised contrastive learning (SupCon) were investigated. The study found that the straightforward supervised approach with BCE loss performed best overall.
  • Explainability: To understand why the models made certain predictions, perturbation-based visual explanations, specifically occlusion sensitivity, were employed. This technique involves occluding parts of the sulci and observing the impact on prediction accuracy, highlighting important regions.

Datasets

The study utilized pooled clinical datasets from three psychiatric disorders:

  • Schizophrenia (SCZ): Data from the schizconnect-vip-prague, cnp, candi, and bsnip1 studies.
  • Bipolar Disorder (BD): Data from the cnp, candi, and bsnip1 studies.
  • Autism Spectrum Disorder (ASD): Data from the abide 1 & 2 studies.

Each dataset included both healthy controls (HC) and patients (PT). The paper notes that these are multicentric datasets, increasing the generalizability of the findings.

Key Insights and Ideas

  • Sulci as Neurodevelopmental Proxies: The study strongly supports the idea that cortical sulci, due to their early developmental origins, are valuable indicators of neurodevelopmental contributions to mental disorders.
  • Simplicity Wins: Surprisingly, the most straightforward supervised learning approach using binary sulcus skeletons outperformed more complex methods like weakly supervised contrastive learning, suggesting that the inherent information in sulcal shapes is potent enough without extensive data augmentation or complex loss functions.
  • Explainability Reveals Potential Biomarkers: The use of explainability techniques like occlusion sensitivity successfully highlighted specific sulcal regions that are important for diagnosis, offering a path to discovering novel biomarkers.
  • Diagnostic Difficulty Varies: The study observed that predicting schizophrenia (SCZ) vs. HC was the most challenging task, followed by bipolar disorder (BD) vs. HC, and then autism spectrum disorder (ASD) vs. HC.

Key Results

  • All trained models demonstrated the ability to predict diagnosis with a ROC AUC greater than 0.57, indicating that cortical sulci contain sufficient information for reliable predictions.
  • DenseNet, an advanced CNN architecture, achieved the best performance, with ROC AUCs of 0.661 for BD vs. HC, 0.656 for SCZ vs. HC, and 0.595 for ASD vs. HC.
  • Supervised learning with a binary cross-entropy loss on binary skeletons yielded the best results, outperforming models trained with Gaussian smoothing or using supervised contrastive learning.
  • Explainability maps successfully highlighted specific sulcal regions implicated in mental disorders, demonstrating the potential for biomarker discovery.

Limitations and Future Work

  • Information Content: The authors acknowledge that sulcus skeleton images might contain less information compared to other sMRI features, which could explain why the prediction performance, while significant, is weaker than some state-of-the-art results.
  • Data Augmentation Suitability: Classical data augmentation techniques may not be well-suited for skeleton images, limiting the benefits of contrastive learning in this context. Future work could involve developing skeleton-specific data augmentations.
  • Patient Subgrouping: The current model classifies patients as a whole group. Future research could explore splitting patients into neurodevelopmental subgroups to refine predictions and investigate finer distinctions.
  • Explainability Refinement: While occlusion sensitivity is useful, it has limitations in capturing long-range dependencies. More advanced explainability techniques, such as SHAP, and more sophisticated patch-based approaches that consider brain atlases, are suggested for future work to better define implicated sulci.
  • Weakly Supervised Pre-training: Leveraging weakly supervised contrastive pre-training on large datasets of healthy controls is proposed as a way to improve performance and stability.

Summary updated

Abstract

Recent advances in machine learning applied to structural magnetic resonance imaging (sMRI) may highlight abnormalities in brain anatomy associated with mental disorders. These disorders are multifactorial, resulting from a complex combination of neurodevelopmental and environmental factors. In particular, such factors are present in cortical sulci, whose shapes are determined very early in brain development and are a valuable proxy for capturing specifically the neurodevelopmental contribution of brain anatomy. This paper explores whether the shapes of cortical sulci can be used for diagnosis prediction using deep learning models. These models are applied to three mental disorders (autism spectrum disorder, bipolar disorder, and schizophrenia) in large multicentric datasets. We demonstrate that the neurodevelopmental underpinnings of these disorders can be captured with sMRI. Finally, we show the potential of visual explanations of models’ decisions in discovering biomarkers for mental disorders.
BibTeX
@inproceedings{auriau2024supervised,
  author = {Pierre Auriau and Antoine Grigis and Benoit Dufumier and Robin Louiset and Joël Chavas and Pietro Gori and Jean-François Mangin and Edouard Duchesnay},
  title = {{Supervised Diagnosis Prediction from Cortical Sulci: Toward the Discovery of Neurodevelopmental Biomarkers in Mental Disorders}},
  booktitle = {ISBI},
  year = {2024},
  doi = {10.1109/isbi56570.2024.10635738},
}
Figure