A deep learning framework for pancreas segmentation with multi-atlas registration and 3D level-set.

Area of Science:

• Medical image analysis
• Deep learning for medical imaging
• Computational anatomy

Background:

• Accurate pancreas segmentation in CT images is crucial for diagnosis and treatment planning.
• Existing segmentation methods often struggle with variations in image quality and anatomical structures.
• Deep learning offers potential for improved segmentation accuracy and robustness.

Purpose of the Study:

• To propose and validate a novel deep learning framework for automated pancreas segmentation in CT volume images.
• To integrate multi-atlas registration and level-set methods within a cascaded coarse-fine-refine pipeline.
• To enhance segmentation accuracy by combining global location, contextual, shape, and edge information.

Main Methods:

• A three-stage segmentation pipeline: coarse (multi-atlas registration), fine (3D and 2D CNNs), and refine (3D level-set).
• Utilized 3D patch-based and 2D slice-based convolutional neural networks (CNNs) for feature learning.
• Incorporated 3D diffeomorphic registration and 3D level-set methods for segmentation refinement.

Main Results:

• The proposed framework achieved an average Dice score exceeding 82% across three diverse CT datasets.
• Demonstrated superior or comparable performance against existing state-of-the-art pancreas segmentation algorithms.
• Validated the framework's effectiveness on datasets with 36, 82, and 281 CT volume images.

Conclusions:

• The cascaded coarse-fine-refine deep learning framework effectively segments the pancreas from CT images.
• The integration of multi-atlas registration, CNNs, and level-set methods enhances segmentation accuracy.
• This approach shows significant promise for clinical applications requiring precise pancreas delineation.