Label refinement network from synthetic error augmentation for medical image segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep convolutional neural networks (CNNs) for medical image segmentation often fail to explicitly learn label structures.
• This can lead to incorrect segmentations, such as disconnected structures in tree-like anatomical features like airways and blood vessels.

Purpose of the Study:

• To propose a novel label refinement method for correcting structural errors in initial image segmentations.
• To implicitly incorporate information about label structure into the segmentation refinement process.

Main Methods:

• A two-part novel method: (1) a model generating synthetic structural errors and (2) a label appearance simulation network creating segmentations with these synthetic errors.
• Training a label refinement network using these synthetic error segmentations and original images to correct and improve initial segmentations.

Main Results:

• The proposed method significantly outperformed a standard 3D U-Net, four previous label refinement methods, and a specialized U-Net for tubular structures on airway and brain vessel segmentation tasks.
• Further improvements were observed when using additional unlabeled data for training.
• An ablation study confirmed the value of individual components of the proposed method.

Conclusions:

• The novel label refinement method effectively corrects structural errors in medical image segmentation.
• The approach shows significant improvements in segmenting complex anatomical structures like airways and blood vessels.
• The method offers a promising solution for enhancing the accuracy and reliability of deep learning-based medical image segmentation.