Transfer Learning for Brain Segmentation: Pre-task Selection and Data Limitations

Manual segmentations of anatomical regions in the brain are time consuming and costly to acquire. In a clinical trial setting, this is prohibitive and automated methods are needed for routine application. We propose a deep-learning architecture that automatically delineates sub-cortical regions in the brain (example biomarkers for monitoring the development of Huntington's disease). Neural networks, despite typically reaching state-of-the-art performance, are sensitive to differing scanner protocols and pre-processing methods. To address this challenge, one can pre-train a model on an existing data set and then fine-tune this model using a small amount of labelled data from the target domain. This work investigates the impact of the pre-training task and the amount of data required via a systematic study. We show that use of just a few samples from the same task (but a different domain) can achieve state-of-the-art performance. Further, this pre-training task utilises automated labels, meaning the pipeline requires very few manually segmented data points. On the other hand, using a different task for pre-training is shown to be less successful. We then conclude, by showing that, whilst fine-tuning is very powerful for a specific data distribution, models developed in this fashion are considerably more fragile when used on completely unseen data.