Leveraging global binary masks for structure segmentation in medical images

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Deep learning (DL) models for medical image segmentation struggle with intensity variations and require extensive training data.
• Existing models often rely solely on pixel intensity, limiting generalization and robustness.

Purpose of the Study:

• To develop a novel framework for organ segmentation in medical images that utilizes anatomical position and shape information.
• To enhance the robustness of DL models to image intensity variations and mitigate challenges associated with limited training data.

Main Methods:

• Proposed a framework using global binary masks to encode recurring anatomical patterns for organ segmentation.
• Evaluated two scenarios: (1) U-Net model trained solely on global binary masks, and (2) global binary masks incorporated as an additional input channel.
• Utilized brain and heart computed tomography (CT) datasets, assessing performance with full and reduced training data subsets.

Main Results:

• Scenario (1) achieved Dice scores of 0.77 ± 0.06 (brain) and 0.85 ± 0.04 (heart), with average Euclidean distances of 3.12 ± 1.43 mm and 2.5 ± 0.93 mm, respectively.
• Scenario (2) significantly improved accuracy (4.3%-125.3% for brain, 1.3%-48.1% for heart) when trained on limited data compared to models using only CT images.
• Global binary masks effectively encoded anatomical position and shape information.

Conclusions:

• The proposed framework demonstrates the potential of leveraging global binary masks for robust medical image segmentation.
• Incorporating anatomical patterns via global binary masks is an effective strategy to improve DL model performance, especially with scarce labeled training data.
• This approach offers a promising solution for developing more generalizable and data-efficient medical image segmentation models.