Work

Data Augmentation to Improve Cross-Domain Generalization in Deep Learning MRI Segmentation

Year: 2021
Type: dissertation
Abstract: Semantic segmentation of medical images is an important task with many applications. However, manually delineating 3D images is time-consuming and the demand for automation is high. For many image segmentation tasks, deep learning has provided state-of-the-art results. However, the varying nature of magnetic resonance (MR) images due to factors such as different machine vendors and clinical protocols has been found to lead to domain transfer issues for deep learning segmentation models. This thesis aims to investigate three different data augmentation methods to remedy this domain generalization problem. We first consider a standard data augmentation approach and study the effects of applying stacked image processing functions to the training data. We then study the effects of including weakly labeled training samples from the unseen data domain in the training set. Lastly we study the effects of training a Cycle GAN to transfer labeled training samples to the unseen domain, and including this synthetic data in the training. The experiments are carried out on MRI data from three separate domains, where labeled training samples only exist for one of them. The studied methods are found to increase average DICE scores on the rectum and urinary bladder by 9-14% and 25-40% on the two unseen domains, and 95th percentile Hausdorff distances are decreased by 60-71% and 33-54%. Additionally, combining the use of unlabeled data from unseen domains with extensive image processing is found to further boost segmentation performance.
Language: English
Open Access status: green
OpenAlex ID: W3193386028
Cites: 17
Cited by:
Related to: 10