Few-shot medical image segmentation using a global correlation network with discriminative embedding.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Supervised learning in deep convolutional neural networks (CNNs) for medical imaging demands extensive annotations, which are costly and time-consuming to acquire.
• Clinical settings often present scenarios with limited annotated data, posing challenges for traditional supervised segmentation models.

Purpose of the Study:

• To develop a few-shot medical image segmentation approach that allows models to rapidly generalize to new classes using only a few training examples.
• To address the limitations of data scarcity in medical image annotation for deep learning models.

Main Methods:

• A few-shot image segmentation mechanism was constructed using a deep convolutional network trained episodically.
• An efficient global correlation module was developed to capture spatial correlations between support and query images, integrated into the deep network.
• The feature embedding scheme was enhanced to improve class discrimination, promoting feature clustering for same-class instances and separation for different-class instances.

Main Results:

• The proposed method demonstrated effective generalization to unseen classes in few-shot learning scenarios.
• Experiments on anatomical abdomen images from CT and MRI modalities validated the approach's performance.
• The global correlation module and enhanced embedding scheme contributed to improved segmentation accuracy with limited data.

Conclusions:

• The developed few-shot segmentation approach offers a viable solution for medical imaging tasks with limited annotations.
• This method has the potential to significantly reduce the annotation burden in clinical practice.
• The technique shows promise for improving the efficiency and applicability of deep learning in medical image analysis.