Abstract Transfer learning has been a very active research topic in natural image processing. But few studies have reported notable benefits of transfer learning on medical imaging. In this study, we sought to investigate the transferability of deep artificial neural networks (DNN) in brain decoding, i.e. inferring brain state using fMRI brain response over a short window. Instead of using pretrained models from ImageNet, we trained our base model on a large-scale neuroimaging dataset using graph convolutional networks (GCN). The transferability of learned graph representations were evaluated under different circumstances, including knowledge transfer across cognitive domains, between different groups of subjects, and among different sites using distinct scanning sequences. We observed a significant performance boost via transfer learning either from the same cognitive domain or from other task domains. But the transferability was highly impacted by the scanner site effect. Specifically, for datasets acquired from the same site using the same scanning sequences, using transferred features highly improved the decoding performance. By contrast, the transferability of representations highly decreased between different sites, with the performance boost reducing from 20% down to 7% for the Motor task and decreasing from 15% to 5% for Working-memory tasks. Our results indicate that in contrast to natural images, the scanning condition, instead of task domain, has a larger impact on feature transfer for medical imaging. With other advanced tools such as layer-wise fine-tuning, the decoding performance can be further improved through learning more site-specific high-level features while retaining the transferred low-level representations of brain dynamics.

Transferability of Brain decoding using Graph Convolutional Networks