Abstract
Deep learning models for Autism Spectrum Disorder (ASD) diagnosis often function as "black boxes," hindering the discovery of neural biomarkers. This paper introduces Invertible Networks (InvNets) to transform opaque classification decisions into interpretable brain connectivity maps. By leveraging the mathematical properties of invertible architectures, the authors map high-dimensional feature-space decision boundaries back to the original input domain. This approach identifies specific brain regions as reliable biomarkers, achieving 71% classification accuracy on the ABIDE dataset and demonstrating superior performance in predicting clinical symptom severity compared to standard methods.
Introduction
The primary challenge in applying deep learning to ASD is interpretability. ASD is typically studied using functional MRI (fMRI) data, represented as connectivity matrices showing interactions between brain regions. While deep neural networks excel at finding patterns in these complex matrices, they are notoriously difficult to interrogate. Existing "explanation" methods provide only rough approximations, which is a major hurdle for clinical adoption. Clinicians require an understanding of which brain circuits are affected to develop targeted treatments and understand the disorder's underlying mechanisms.
Methods
The researchers proposed a novel classification pipeline based on Invertible Blocks. Unlike standard networks where information is lost across layers, an invertible network is designed so the input can be perfectly reconstructed from the output.
The architecture consists of several invertible blocks followed by a single linear Fully-Connected (FC) layer. Because the final layer is linear, the decision boundary is a simple plane in the high-dimensional feature space. By "inverting" the network, the authors can map points from this decision boundary back to the original input domain (the brain connectivity matrix). This allows for the calculation of an "importance" score for every neural connection, ranking pathways based on their criticality to the diagnosis.
Experimental Setup
The study utilized the Autism Brain Imaging Data Exchange (ABIDE) dataset, comprising 530 subjects with ASD and 505 control subjects. Data was processed using the CCS pipeline and mapped to the CC200 atlas, dividing the brain into 200 regions of interest. This resulted in a connectivity vector of 19,900 unique edges per subject. To improve robustness, bootstrapping was used to generate over 51,000 training examples.
Results and Discussion
The InvNet achieved 71% accuracy on the ABIDE dataset, outperforming Support Vector Machines (67%), Random Forests (66%), and standard Deep Neural Networks (70%).
A core insight of this work is that while the relationship between brain connectivity and ASD is non-linear, the InvNet learns a transformation that linearizes this relationship in a hidden feature domain. Mapping the linear boundary back to the input domain reveals the complex, non-linear boundary in the original space.
The model identified biomarkers in regions associated with social perception and executive function, such as the superior temporal gyrus, frontal cortex, and insular cortex. Crucially, using only the top 10% of "important" edges identified by the model to predict clinical severity scores (SRS) resulted in lower error rates than using the full connectivity dataset. This suggests the method effectively filters noise to isolate clinically relevant signals.
Conclusion
While the approach is currently dependent on specific brain atlases and involves computational complexities inherent to invertible architectures, it provides a rigorous framework for biomarker discovery. By turning the "black box" into an invertible map, this research bridges the gap between high-performance deep learning and clinical interpretability in neuroimaging.
Summary updated