Pairwise learning for medical image segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Fully convolutional networks (FCNs) are effective for medical image segmentation but struggle with limited labeled data and appearance variability.
• Overfitting is a significant challenge due to intra-class heterogeneity and boundary ambiguity in small datasets.

Purpose of the Study:

• To develop an improved FCN-based framework for medical image segmentation that addresses data scarcity and appearance variability.
• To introduce a novel approach, proxy supervision, to leverage prior label space information for enhanced segmentation accuracy.

Main Methods:

• Proposes a conjugate fully convolutional network (CFCN) that processes pairwise samples for rich context representation via a fusion module.
• Introduces proxy supervision to exploit label space priors, mitigating overfitting with limited training data.
• Develops a compact conjugate fully convolutional network (C²FCN) with a single head for efficient proxy supervision fitting.

Main Results:

• The proposed CFCN and C²FCN frameworks demonstrate significant performance improvements in both binary and multi-category medical image segmentation.
• The methods show particular effectiveness when dealing with limited amounts of training data.
• Evaluations on the Liver Tumor Segmentation (LiTS) and Combined (CT-MR) Healthy Abdominal Organ Segmentation (CHAOS) datasets validate the framework's efficacy.

Conclusions:

• The conjugate fully convolutional network (CFCN) and its compact variant (C²FCN) offer a robust solution for medical image segmentation, especially under data-scarce conditions.
• Proxy supervision is an effective strategy for improving segmentation performance by leveraging label space information.
• The developed framework provides a valuable tool for enhancing the accuracy and efficiency of medical image analysis.