Methods for 2-D and 3-D Endobronchial Ultrasound Image Segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Pulmonology

Background:

• Endobronchial ultrasound (EBUS) is crucial for lung cancer staging during bronchoscopy.
• Interpreting EBUS video sequences is challenging, hindering its widespread adoption.
• Computer-based image segmentation has advanced other ultrasound fields but not EBUS.

Purpose of the Study:

• To develop and evaluate automated image segmentation methods for 2-D EBUS frames and 3-D EBUS sequences.
• To address the challenges in EBUS interpretation and improve its utility in clinical practice.
• To demonstrate the potential of these methods for EBUS localization in image-guided bronchoscopy.

Main Methods:

• Proposed novel image segmentation methods for 2-D EBUS frames using fast-marching level-set, anisotropic diffusion, and region growing.
• Developed a 3-D segmentation method building upon the 2-D approach, incorporating geodesic level-set for EBUS sequences.
• Tested methods on lung cancer patient data, assessing automation levels, accuracy (Dice index), and computation time.

Main Results:

• Methods achieved full automation in nearly 80% of cases, with minimal user interaction (seed point selection) in others.
• The 2-D method yielded a Dice index of 90.0% ±4.9%, and the 3-D method achieved 83.9% ± 6.0% compared to ground truth.
• Computation times were significantly faster (0.070 s/frame for 2D, 0.088 s/frame for 3D) than manual contouring.

Conclusions:

• Developed effective, automated image segmentation techniques for 2-D and 3-D EBUS data.
• These methods offer a substantial improvement in speed and accuracy for EBUS interpretation.
• The proposed techniques show promise for enhancing EBUS localization within multimodal image-guided bronchoscopy systems.