Extracting Interpretable Brain Biomarkers from Deep Learning Models Trained on rs-fMRI Data

Mental health issues, particularly conditions such as Autism Spectrum Disorder (ASD), are a major global concern. The lack of reliable biomarkers poses a challenge for timely and accurate diagnosis, which relies solely on behavioral observations. This study explores the intersection of deep learning, explicable AI, and ASD diagnosis, with the goal of extracting meaningful biomarkers from brain imaging data. The present study complements an earlier study in which we developed a predictive modelling approach for ASD.

We focus on evaluating methods that help interpret machine learning predictions, using a model called METAFormer trained on ASD data. Our goal is to identify the best method for explaining the model's decisions and improve our understanding of how it works. Beyond prediction, we are exploring brain imaging-based biomarkers to deepen our understanding of ASD.

We use a variety of attribution methods to identify critical regions of interest (ROIs) in the brain. We evaluate these methods using metrics such as infidelity and sensitivity, which measure the accuracy of explanations under significant perturbations and how attribution is affected by small changes, respectively.

Our results show that DeepLIFT strikes the best balance between accuracy and sensitivity. Choosing an appropriate baseline value is critical, and we find that a baseline of -1 provides robust and faithful explanations. When examining important features, our results are consistent with the existing literature, highlighting the role of specific brain regions in discriminating between ASD and controls.

In conclusion, DeepLIFT emerges as the most effective method for extracting ROIs from our model, providing consistent and literature-aligned explanations. The identified brain regions may be critical for the development of deep learning-based imaging biomarkers, paving the way for more accurate diagnosis in mental health. This study opens avenues for further research in deep learning-based biomarker discovery.