Continual Hippocampus Segmentation with Transformers

TL;DR

This paper investigates the use of Transformer-based models for continual hippocampus segmentation, demonstrating their potential to reduce catastrophic forgetting in medical image analysis compared to traditional convolutional models.

Contribution

It provides an analysis of Transformer mechanisms in sequential learning for medical segmentation and offers insights on adapting continual learning strategies for these models.

Findings

Transformers mitigate catastrophic forgetting better than convolutional models.

Regularising ViT modules requires careful consideration.

Transformer mechanisms improve robustness in sequential hippocampus segmentation.

Abstract

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.