Active learning based segmentation of Crohns disease from abdominal MRI

Area of Science:

• Medical imaging analysis
• Machine learning in healthcare
• Gastroenterology

Background:

• Accurate segmentation of Crohn's disease (CD) tissues in abdominal magnetic resonance (MR) images is crucial for diagnosis and treatment monitoring.
• Fully supervised learning (FSL) methods demand extensive labeled datasets, which are costly and require specialized expertise to acquire.
• Semi-supervised learning (SSL) and active learning (AL) offer potential solutions by leveraging unlabeled data and optimizing the labeling process.

Purpose of the Study:

• To develop and evaluate a novel active learning (AL) framework integrated with semi-supervised learning (SSL) for enhanced segmentation of Crohn's disease (CD) tissues in abdominal MR images.
• To reduce the dependency on large, expert-annotated datasets for robust medical image segmentation.
• To improve the efficiency and accuracy of CD tissue segmentation compared to traditional FSL methods.

Main Methods:

• A new AL framework was designed, incorporating a query strategy that combines classification uncertainty, feature similarity, and novel context information.
• This AL framework was integrated with SSL to leverage a small set of labeled MR images and a large pool of unlabeled images.
• The combined SSL-AL approach was applied to segment Crohn's disease tissues in abdominal MR images.

Main Results:

• The proposed SSL-AL method achieved higher segmentation accuracy compared to FSL methods when using fewer labeled samples.
• The framework demonstrated a significant reduction in overall training time and labeling effort.
• Experimental results validated the effectiveness of the combined approach in optimally utilizing limited labeled data and abundant unlabeled data.

Conclusions:

• The integrated SSL-AL framework provides a robust and efficient method for segmenting Crohn's disease tissues in abdominal MR images.
• This approach significantly lowers the requirements for labeled data and expert annotation, making advanced segmentation more accessible.
• The novel query strategy enhances the performance of AL by effectively selecting informative samples for labeling, leading to improved segmentation accuracy and reduced training burden.