NeuroExplainer: Fine-Grained Attention Decoding to Uncover Cortical Development Patterns of Preterm Infants

Chenyu Xue, Fan Wang, Yuanzhuo Zhu, Hui Li, Deyu Meng, Dinggang Shen, Chunfeng Lian

Introduction

Deep learning has demonstrated significant potential in medical imaging, particularly in identifying biomarkers for neurodevelopmental disorders. However, the "black box" nature of these models remains a significant hurdle for clinical adoption. In neuroscience, classification accuracy is insufficient on its own; researchers must understand the underlying neurobiological drivers—specifically, which cortical regions deviate from typical development.

Existing interpretability methods typically rely on "post-hoc" explanations, such as heatmaps generated after a model is trained. These methods assume that a high-performing model naturally produces faithful explanations. However, in neuroimaging—where datasets are often small and 3D brain meshes are high-dimensional and noisy—this assumption frequently fails. NeuroExplainer addresses this by integrating explainability directly into the training process, treating the search for explanation factors as a mechanism to improve classification performance rather than an afterthought.

The NeuroExplainer Framework

NeuroExplainer is an explainable geometric deep network designed to operate on spherical representations of the cerebral cortex. By working directly on 3D meshes (non-Euclidean data) rather than 2D slices or 3D voxels, the model preserves the complex, folded topology of the human brain.

Hierarchical Attention-Decoding

The architecture utilizes an encoder-decoder structure tailored for spherical surfaces. The encoder (EB-1 to EB-4) extracts features at progressively lower resolutions using hexagonal convolutions. The decoder (DB-1 to DB-3) then "unpacks" these features using a hierarchical attention mechanism. This allows the model to propagate coarse, high-level insights back up to high-resolution cortical surfaces, resulting in fine-grained, vertex-level attention maps.

Domain-Knowledge-Guided Regularization

The core innovation of NeuroExplainer lies in its three specific constraints based on neurobiological development patterns:

  1. Fidelity-Aware Contrastive Learning: This forces the model to ensure that the "atypical" regions it identifies are the primary drivers of the classification, ensuring the explanation is faithful to the model's decision-making process.
  2. Sparsity-Aware Regularization: Using entropy-based losses, the model is incentivized to identify localized, distinct brain changes rather than scattered, random noise across the cortex.
  3. Stability-Aware Regularization: Through a custom data augmentation strategy involving random mesh coarsening and cross-scale consistency, the model ensures that explanations remain stable even when input data is slightly perturbed.

Experimental Setup and Dataset

The study utilized the developing Human Connectome Project (dHCP) benchmark, a comprehensive dataset of infants scanned at term-equivalent age (35–44 weeks). The cohort included 700 subjects: 143 preterm and 557 term-born infants.

To facilitate geometric deep learning, 3D brain meshes were mapped to spheres. The model utilized three fundamental vertex-wise cortical attributes as input features:

  • Cortical thickness
  • Mean curvature
  • Average convexity

Results and Discussion

NeuroExplainer significantly outperformed state-of-the-art geometric models, including SphericalCNN, SphericalMoNet, and SubdivNet, across multiple performance and interpretability metrics.

Classification Performance

The model achieved a state-of-the-art Accuracy (ACC) of 0.95 and an Area Under the Curve (AUC) of 0.97. Notably, it showed a substantial jump in sensitivity (0.94), indicating a superior ability to correctly identify preterm infants within an imbalanced dataset compared to baseline models (which ranged from 0.79 to 0.93 ACC).

Quantitative and Qualitative Explainability

NeuroExplainer consistently outperformed post-hoc methods like CAM and Grad-CAM in Fidelity, Sparsity, and Stability. It achieved a Stability score of 0.96, nearly double that of several competitors.

Qualitatively, the attention maps generated by the model highlighted regions in the inferior parietal and superior temporal lobes. These findings are highly significant as they align with existing multi-modal neuroimaging literature regarding the impact of preterm birth on cortical development. This suggests that the model is not only accurate but is identifying biologically grounded patterns of change.

Conclusion and Future Directions

NeuroExplainer demonstrates that explainability can serve as a "booster" for classification. By forcing the model to find sparse, stable, and high-fidelity factors, the researchers improved the model's ability to distinguish preterm from term-born infants.

While the results are promising, future work will focus on expanding the feature scope to include functional MRI (fMRI) or diffusion MRI (dMRI) data for a more holistic view of development. Additionally, further clinical validation by neonatologists is required to determine the predictive value of these identified regions for long-term neurodevelopmental outcomes. The framework's "explanation-driven" approach also holds potential for application in adult neurodegenerative diseases, such as Alzheimer’s.

Summary updated

BibTeX
@inproceedings{xue2023neuroexplainer,
  author = {Chenyu Xue and Fan Wang and Yuanzhuo Zhu and Hui Li and Deyu Meng and Dinggang Shen and Chunfeng Lian},
  title = {{NeuroExplainer: Fine-Grained Attention Decoding to Uncover Cortical Development Patterns of Preterm Infants}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2023},
  url = {https://openreview.net/forum?id=GjAqpVsRRu},
}
Figure