Bidirectional teaching between lightweight multi-view networks for intestine segmentation from CT volume.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate intestine segmentation is critical for diagnosing and treating intestinal diseases like obstruction.
• Limited labeled data due to complex spatial structures hinders fully supervised learning in medical image segmentation.
• Existing methods struggle with the scarcity of annotated data for precise intestinal segmentation.

Purpose of the Study:

• To develop a semi-supervised method for accurate intestine segmentation from computed tomography (CT) volumes.
• To enhance segmentation performance by effectively utilizing limited labeled and abundant unlabeled data.
• To address the challenges posed by complex intestinal spatial features in medical image analysis.

Main Methods:

• A 3D segmentation network employing a bidirectional teaching strategy with simultaneously trained backbones.
• Generation of pseudo-labels from unlabeled data to augment the training dataset.
• Implementation of a lightweight multi-view symmetric network with small convolutional kernels for multi-scale feature extraction.

Main Results:

• The proposed semi-supervised method achieved an average Dice score of 80.45% on 59 CT volumes.
• The method demonstrated an average precision of 84.12% and an average recall of 78.84%.
• Experimental results were validated through five repetitions, indicating robust performance.

Conclusions:

• The semi-supervised approach effectively leverages unlabeled data via pseudo-labeling, crucial for medical image segmentation with limited annotations.
• Assigning differential weights to pseudo-labels enhances their reliability and improves overall segmentation accuracy.
• The proposed method offers competitive performance compared to existing techniques for intestine segmentation.