Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography

Area of Science:

• Ophthalmic imaging analysis
• Medical image segmentation
• Deep learning applications in radiology

Background:

• Deep learning models are increasingly used for ophthalmic image segmentation.
• Accurate segmentation of orbital structures in CT scans is crucial for diagnosis and treatment planning.
• Traditional manual segmentation is time-consuming and labor-intensive.

Purpose of the Study:

• To develop and evaluate a U-Net based deep learning model for segmenting orbital computed tomography (CT) images.
• To introduce an efficient labeling method using super-resolution for ground truth masking.
• To segment key orbital structures including the eyeball, optic nerve, and extraocular muscles.

Main Methods:

• Construction of an orbital CT segmentation model using the U-Net architecture.
• Implementation of a super-resolution labeling technique for efficient ground truth generation.
• Dataset pre-processing involving cropping volumes of interest.
• Utilizing sequential 2D slices as input with bi-directional convolutional long-term short memories to capture inter-slice correlations.

Main Results:

• A U-Net model was successfully constructed for orbital CT segmentation.
• The method demonstrated efficient masking of ground truth using super-resolution labeling.
• The model focused on segmenting the eyeball, optic nerve, and extraocular muscles.
• Evaluation indicated the potential of deep learning for orbital CT image segmentation.

Conclusions:

• Deep learning, specifically U-Net, offers a viable approach for orbital CT image segmentation.
• The proposed super-resolution labeling method enhances the efficiency of creating training datasets.
• The developed model shows promise for clinical applications in ophthalmology and orbital imaging.