Unsupervised 3D Lung Segmentation by Leveraging 2D Segment Anything Model.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Lung segmentation is crucial for lung nodule detection and cancer analysis.
• Deep learning excels in medical image analysis but requires extensive annotated data, which is scarce in medical imaging.
• Existing supervised methods for lung segmentation demand precise ground truth annotations.

Purpose of the Study:

• To develop an unsupervised 3D lung segmentation method for CT data.
• To leverage the capabilities of a foundational 2D Segment Anything Model (SAM) for this task.
• To eliminate the need for manual annotations in training 3D lung segmentation models.

Main Methods:

• Utilized a 2D Segment Anything Model (SAM) to generate initial 2D lung masks from individual CT slices.
• Reconstructed 3D lung masks by integrating multiple 2D masks from the same 3D CT scan.
• Trained a 3D lung segmentation model using these reconstructed 3D masks in a fully unsupervised manner.

Main Results:

• The proposed unsupervised 3D lung segmentation model demonstrated performance comparable to supervised methods on the LUNA16 dataset.
• The unsupervised approach exhibited enhanced stability in segmentation results compared to traditional supervised training.
• Successfully generated 3D lung masks without relying on any ground truth annotations.

Conclusions:

• Unsupervised 3D lung segmentation is feasible and effective by leveraging foundational 2D models like SAM.
• This method significantly reduces the dependency on laborious manual annotation processes in medical imaging.
• The developed technique offers a promising alternative for robust and stable lung segmentation in clinical applications.