Semi-supervised and active learning for automatic segmentation of Crohn's disease

Area of Science:

• Medical Imaging
• Machine Learning
• Gastroenterology

Background:

• Crohn's disease (CD) diagnosis relies on accurate interpretation of abdominal magnetic resonance (MR) images.
• Manual segmentation and classification of CD in MR images are time-consuming and require expert knowledge.
• Existing automated methods may require extensive labeled data, limiting their practical application.

Purpose of the Study:

• To develop and evaluate a novel method for automatic detection and segmentation of Crohn's disease in abdominal MR images.
• To combine the strengths of semi-supervised learning (SSL) and active learning (AL) for improved efficiency and accuracy.
• To utilize Random Forest (RF) classifiers for robust and interpretable analysis.

Main Methods:

• A hybrid approach integrating SSL and AL was developed for CD detection and segmentation.
• Random Forest (RF) classifiers were employed for efficient SSL classification and knowledge interpretation.
• A novel information density weighted approach, incorporating context, semantic knowledge, and labeling uncertainty, was used for active learning sample selection.

Main Results:

• The proposed method effectively combined the advantages of both SSL and AL.
• Higher classification and segmentation accuracy for Crohn's disease was achieved compared to fully supervised methods.
• The approach demonstrated superior performance with a significantly reduced number of labeled samples.

Conclusions:

• The integrated SSL and AL method offers a promising solution for automated Crohn's disease analysis in MR imaging.
• This approach enhances diagnostic efficiency by reducing the need for extensive manual labeling.
• The findings suggest potential for improved clinical workflows in managing Crohn's disease.