Deep Learning Guided Partitioned Shape Model for Anterior Visual Pathway Segmentation.

Area of Science:

• Medical Imaging
• Neuroscience
• Computer Vision

Background:

• Segmenting the anterior visual pathway (AVP) in MRI is difficult due to its thin structure, variability, and low contrast.
• Pathological AVP segmentation, e.g., for gliomas, presents additional segmentation challenges.

Purpose of the Study:

• To develop a fully automated partitioned shape model for AVP segmentation using multi-sequence MRI and deep learning.
• To create a joint statistical shape model capable of segmenting both healthy and pathological AVPs.

Main Methods:

• A novel framework combining deep learning features with a partitioned statistical shape model.
• Utilized conditional space deep learning for robust shape localization.
• Employed volumetric multiscale curvelet transform for intensity normalization.
• Developed optimally partitioned statistical shape and appearance models for regional flexibility.

Main Results:

• Achieved a mean Dice similarity coefficient of 0.779 for entire AVP segmentation and 0.791 for the optic nerve.
• Demonstrated superior performance compared to state-of-the-art methods.
• Showed robustness comparable to manual segmentation in pediatric subjects.

Conclusions:

• The proposed automated segmentation method accurately segments the AVP in multi-sequence MRI.
• Deep learning significantly aids segmentation in low-contrast and pathological regions.
• This approach offers a robust and efficient tool for AVP analysis in clinical research.