Clarify mechanisms of transfer learning and quantify deep ensemble benefits in psychiatric MRI classification

Determine the mechanisms by which transfer learning improves predictive performance and quantify the magnitude of performance improvements achieved by deep ensemble averaging in single-subject classification of bipolar disorder and schizophrenia from 3D whole-brain anatomical MRI data.

Background

Prior work reported that transfer learning and deep ensembles can outperform traditional machine learning and randomly initialized deep networks for single-subject classification of psychiatric disorders using 3D whole-brain anatomical MRI. However, the specific reasons for these gains and the extent to which deep ensembles contribute to performance had not been clearly established. This paper sets out to investigate these questions by examining training stability, ensemble size effects, and loss landscape properties for bipolar disorder and schizophrenia classification.

The study compares transfer learning models—initialized from an age-aware contrastive pretraining on healthy brains—with randomly initialized deep learning models. It analyzes how many models are needed in an ensemble to achieve substantial performance improvements and explores whether transfer learning constrains fine-tuned models to the same basin in the loss landscape, potentially explaining improved robustness and generalization.