Multi-dimensional consistency learning between 2D Swin U-Net and 3D U-Net for intestine segmentation from CT volume

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate intestine segmentation in CT volumes is challenging due to complex anatomical structures and limited labeled datasets.
• Existing methods struggle with precise pixel-level labeling of intestines in abdominal CT scans.

Purpose of the Study:

• To introduce a novel two-step network employing semi-supervised learning for improved intestine segmentation from CT volumes.
• To address the difficulties in segmenting complex intestinal structures and overcome limitations of small labeled datasets.

Main Methods:

• A two-stage semi-supervised learning model was developed, combining 2D Swin U-Net and 3D U-Net.
• Stage 1 involves generating pseudo-labels for unlabeled data using a 2D Swin U-Net trained on labeled data.
• Stage 2 utilizes a 3D U-Net trained on both labeled and pseudo-labeled data for final segmentation.

Main Results:

• The proposed method demonstrated improved performance on 59 CT volumes.
• Achieved an average increase of 3.25% in Dice score and 6.84% in recall rate compared to baseline methods.
• The multi-dimensional consistency learning approach effectively mitigated segmentation inaccuracies.

Conclusions:

• The novel semi-supervised learning method, integrating 2D and 3D U-Nets, significantly enhances intestine segmentation accuracy.
• The approach effectively leverages limited labeled data and maintains multi-dimensional output consistency for superior results.
• This method offers a promising solution for accurate intestine segmentation in medical imaging applications.