Enhancement of the Domain Generalization of Vision Transformers through Advanced Data Augmentation Techniques

Abstract

Domain generalization is a critical challenge in medical imaging, where the performance of deep learning models can significantly degrade due to domain shifts—variations in data distribution caused by differences in imaging protocols, equipment, or patient populations. This issue is particularly problematic in medical image segmentation, where models trained on a specific dataset often fail to generalize to new, unseen data, limiting their practical applicability in clinical settings. Addressing this problem requires innovative techniques to enhance model robustness and generalizability. This thesis investigates the enhancement of domain generalization in medical image segmentation through advanced data augmentation techniques. This study focuses on evaluating the impact of data augmentation at both the input and feature levels using methods such as style-based augmentation. By incorporating these augmentation strategies, the goal is to improve the ability of models to handle unseen variations in medical images, thereby increasing their robustness and reliability in real-world applications. In the experiments, a vision transformer model was fine-tuned with datasets augmented through a combination of style-based and other input-level augmentation methods. These techniques enhance the diversity of training data, allowing the model to learn robust features that are less sensitive to various types of data shifts. The evaluation was conducted on prostate MRI datasets as the in-domain data and six additional datasets as the out-of-distribution domains. The results demonstrated that models trained with augmented data exhibited significantly improved robustness and generalization to OOD samples. The combination of style- based and other augmentation methods led to a notable increase in generalizability. This suggests that integrating complex data augmentation techniques can significantly enhance the robustness of medical image segmentation models, making them more reliable for clinical applications.