Analysis of Whole-Brain Resting-State MRI Using Multi-Label Deformable Offset Networks and Segmentations Based Attention with Explorations into the Ethical Implications of Artificial Intelligence in Clinical Psychiatry Settings and Care

Abstract

Due to the poor understanding of the underlying biological mechanisms of psychiatric disorders, diagnoses rely upon symptomatic criteria and clinicians’ discretion. Reviews of these criteria have revealed issues of heterogeneity, over and under specificity, and symptom overlap between disorders. Deep learning provides a method to produce quantifiable diagnostic labels based upon biological markers such as specific features of brain anatomy or functionality. In practice, these methods fail to indicate how a particular result was determined, raising major obstacles for clinical implementation.To improve the efficiency and interpretability of existing deep networks, we have developed a novel atlas-based attention module to more easily capture global information across different areas of brain function. Our model can be extended to symptom level classification using NIMH data to give clinicians usable information outside of broad disorder classification. We have compared our model against leading 3D deep learning frameworks and have shown that our novel atlas-based attention module achieves 88% F1 and 91% accuracy on the UCLA Consortium for Neuropsychiatric Phenomics dataset. We have embedded our model with elements like deformable convolutions, gradient activation visualizations, and occlusion testing to show model attention and function. In addition to the lack of explainability, addressing the ethical issues surrounding clinical implementation of artificial intelligence is necessary before usage can become a reality. We identified a series of regulatory recommendations to address pertinent ethical concerns of equity and bias during both model development and clinical usage. We propose a standardized protocol for developing a clinical reference standard, the development of diversity reports regarding data used by models, and regulation of usage scenarios to reduce contextual bias.