Medical Image Segmentation Using Transformer Networks.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep learning, particularly fully-convolutional networks (FCNs), dominates medical image segmentation.
• FCNs utilize convolution operations, offering benefits like parameter sharing but struggling with long-range interactions and content-agnostic processing.
• Limitations of FCNs include fixed operations and difficulty modeling long-range dependencies in medical images.

Purpose of the Study:

• To introduce a novel deep neural network architecture for medical image segmentation that overcomes FCN limitations.
• To demonstrate the efficacy of a transformer-based, self-attention mechanism for segmentation tasks.
• To address the challenge of limited labeled medical data through effective pre-training strategies.

Main Methods:

• A transformer network architecture is proposed, processing 3D image blocks by dividing them into patches.
• Self-attention mechanisms are computed between 1D embeddings of neighboring image patches for segmentation prediction.
• Methods for pre-training the model on large unlabeled medical image datasets are developed to mitigate data scarcity.

Main Results:

• The proposed self-attention-based model achieved higher segmentation accuracies than state-of-the-art FCNs on two benchmark datasets.
• The network demonstrated strong performance when trained with very few labeled images (tens).
• Pre-training strategies significantly enhanced the model's performance, especially in low-data regimes, outperforming FCNs.

Conclusions:

• A novel, convolution-free, transformer-based architecture offers a more accurate approach to medical image segmentation.
• The model's reliance on self-attention effectively captures necessary image context, surpassing FCN capabilities.
• The proposed pre-training methods make the model highly effective even with scarce labeled medical imaging data.