Continual Hippocampus Segmentation with Transformers
This work addresses the need for robust segmentation models in radiology to handle evolving patient data, though it is incremental as it adapts existing methods to a new domain.
The paper tackled the problem of catastrophic forgetting in medical image segmentation for changing clinical conditions by evaluating Transformer-based architectures in continual learning scenarios, finding that they mitigate forgetting compared to convolutional methods and require careful regularization of ViT modules.
In clinical settings, where acquisition conditions and patient populations change over time, continual learning is key for ensuring the safe use of deep neural networks. Yet most existing work focuses on convolutional architectures and image classification. Instead, radiologists prefer to work with segmentation models that outline specific regions-of-interest, for which Transformer-based architectures are gaining traction. The self-attention mechanism of Transformers could potentially mitigate catastrophic forgetting, opening the way for more robust medical image segmentation. In this work, we explore how recently-proposed Transformer mechanisms for semantic segmentation behave in sequential learning scenarios, and analyse how best to adapt continual learning strategies for this setting. Our evaluation on hippocampus segmentation shows that Transformer mechanisms mitigate catastrophic forgetting for medical image segmentation compared to purely convolutional architectures, and demonstrates that regularising ViT modules should be done with caution.