# Promises and pitfalls of deep neural networks in neuroimaging-based psychiatric research

**Authors:** Fabian Eitel, Marc-André Schulz, Moritz Seiler, Henrik Walter, Kerstin Ritter  
**arXiv:** [2301.08525](https://arxiv.org/abs/2301.08525)  
**Date:** 2023-01-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2301.08525v1/

## Key points

- Deep neural networks are a powerful tool in neuroimaging-based psychiatric research for accurate diagnostics and treatment recommendations.
- Methodological concepts such as representation and transfer learning, as well as domain-specific priors, contribute to the advancement of deep neural networks in this field.
- Applications include diagnosing psychiatric diseases, delineating disease subtypes, normative modeling, and developing neuroimaging biomarkers.
- Challenges include training models on small, heterogeneous, and biased datasets, lack of validity in clinical labels leading to potential bias, and confounding variables affecting result interpretation.

## Summary

In the field of neuroimaging-based psychiatric research, deep neural networks have emerged as a powerful tool for providing more accurate diagnostics and individual treatment recommendations. This advancement is attributed to key methodological concepts such as representation and transfer learning, as well as the modeling of domain-specific priors. Recent applications in this area include the diagnosis of psychiatric diseases, delineation of disease subtypes, normative modeling, and the development of neuroimaging biomarkers. However, despite these promising developments, there are several challenges that need to be addressed. One major issue is the difficulty of training models on small, heterogeneous, and biased datasets. Additionally, there is a lack of validity in clinical labels used for training these models, leading to potential algorithmic bias. Moreover, the influence of confounding variables poses a significant challenge in accurately interpreting results obtained from deep neural networks in neuroimaging-based psychiatric research. Their work provides valuable insights into the current state of utilizing deep neural networks for medical imaging applications within psychiatric research while highlighting areas that require further attention and refinement for future advancements in this field.