3D Vascular Segmentation Supervised by 2D Annotation of Maximum Intensity Projection

Area of Science:

• Medical Image Analysis
• Computational Biology
• Machine Learning in Healthcare

Background:

• Accurate vascular structure segmentation is vital for medical analysis but hindered by laborious 3D annotation requirements for fully supervised models.
• Existing weakly-supervised methods struggle with sparse vascular structures, necessitating improved annotation strategies.
• Maximum Intensity Projection (MIP) offers a potential solution for dimensionality reduction and efficient annotation.

Purpose of the Study:

• To develop a novel weakly-supervised approach for 3D vascular segmentation that overcomes limitations of existing methods.
• To leverage 2D MIP projections for efficient vascular annotation and guide 3D segmentation model training.
• To reduce the time and effort associated with manual vessel annotation in medical imaging.

Main Methods:

• Utilized Maximum Intensity Projection (MIP) to reduce 3D volumes to 2D images for efficient annotation.
• Generated pseudo-labels for 3D vessels from 2D projection annotations.
• Developed a weakly-supervised network fusing 2D-3D deep features via MIP, incorporating confidence learning and uncertainty estimation for pseudo-label refinement and model fine-tuning.

Main Results:

• The proposed method achieved highly competitive performance in segmenting various vascular structures across five diverse datasets (cerebral vessels, aorta, coronary artery).
• Demonstrated significant potential in reducing the time and effort required for vessel annotation.
• Validated the effectiveness of fusing 2D-3D deep features through MIP for improved segmentation accuracy.

Conclusions:

• The developed weakly-supervised method effectively addresses the challenges of 3D vascular segmentation by utilizing 2D MIP annotations.
• This approach offers a practical and efficient solution for medical image analysis, significantly reducing annotation burden.
• The technique shows promise for broader clinical applications requiring precise vascular segmentation.