Sd-net: a semi-supervised double-cooperative network for liver segmentation from computed tomography (CT) images

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate liver segmentation is vital for quantitative biomarker extraction in clinical diagnosis and computer-aided decision support.
• Challenges in liver segmentation include noise, artifacts in CT images, complex backgrounds, and indistinct boundaries.
• Fully supervised methods require extensive manual annotations, which are labor-intensive and require expert knowledge, leading to limited high-quality datasets.

Purpose of the Study:

• To develop a semi-supervised method for liver segmentation that overcomes the limitations of fully supervised approaches.
• To leverage both limited labeled and large unlabeled datasets for improved segmentation accuracy.
• To enhance the efficiency and reduce the reliance on expert manual annotation in medical image segmentation.

Main Methods:

• A semi-supervised double-cooperative network (SD-Net) was developed, comprising two collaborative network models.
• An adaptive mask refinement approach was introduced in the supervised module for precise segmentation from labeled data.
• A dynamic pseudo-label generation strategy was employed in the unsupervised module to utilize unlabeled data effectively.

Main Results:

• The proposed SD-Net method achieved a Dice score exceeding 94% for liver segmentation.
• The method demonstrated high accuracy in segmenting liver volumes from preoperative abdominal CT images.
• Experimental findings indicate the robustness of the approach in handling challenging image characteristics.

Conclusions:

• The developed SD-Net offers a highly accurate and efficient solution for automatic liver segmentation.
• The semi-supervised approach effectively utilizes weakly labeled data, reducing annotation burden.
• The method shows significant potential for integration into routine clinical practice for improved diagnostic accuracy.