Intra- and Inter-Pair Consistency for Semi-Supervised Gland Segmentation.

Area of Science:

• Digital Pathology
• Medical Image Analysis
• Computer Vision

Background:

• Accurate gland segmentation in histology images is crucial for disease diagnosis but hindered by the scarcity of annotated data.
• Deep learning models excel at segmentation but typically require extensive labeled datasets, posing a significant challenge.

Purpose of the Study:

• To develop a semi-supervised model for gland segmentation that effectively utilizes both labeled and unlabeled histology images.
• To address the data annotation bottleneck in medical image segmentation tasks.

Main Methods:

• Proposed an intra- and inter-pair consistency-based semi-supervised (I²CS) model for gland segmentation.
• Introduced a novel consistency module to learn from unlabeled data by exploring semantic consistency across image pairs.
• Designed an object-level loss (Obj-Dice loss) to handle challenges posed by touching glands.

Main Results:

• The I²CS model demonstrated superior performance compared to recent gland segmentation and semi-supervised methods on the GlaS and CRAG datasets.
• The proposed consistency module and Obj-Dice loss were shown to be effective in improving segmentation accuracy.
• Achieved state-of-the-art gland segmentation performance on benchmark datasets.

Conclusions:

• The I²CS model offers an effective solution for gland segmentation, significantly reducing reliance on large annotated datasets.
• The developed methods provide a robust framework for semi-supervised learning in medical image analysis.
• The findings highlight the potential of consistency-based learning for advancing digital pathology.