Lens structure segmentation from AS-OCT images via shape-based learning.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Accurate segmentation of ocular structures like the lens nucleus and cortex from AS-OCT is crucial for diagnosing conditions such as cataracts and presbyopia.
• Weak-contrast boundaries in AS-OCT images pose a significant challenge for current segmentation methods, including U-Net, which treat segmentation as a pixel classification task.

Purpose of the Study:

• To develop an improved deep learning framework for accurate segmentation of the lens nucleus and cortex in AS-OCT images.
• To address the limitations of existing methods in handling weak-contrast boundaries by incorporating shape prior information.

Main Methods:

• Proposed a deep learning framework that learns a level set function, with its zero-level set representing the object boundary, using convolutional neural networks.
• Introduced a novel shape-based loss function to naturally embed shape prior knowledge into the learning process.
• Trained the model on a high-quality AS-OCT dataset with precise annotations.

Main Results:

• Achieved high segmentation performance with mean Intersection over Unions (MIoUs) of 0.946 for the nucleus and 0.957 for the cortex.
• Demonstrated superior boundary accuracy with mean Euclidean Distance (MED) of 6.746 pixels for the nucleus and 2.045 pixels for the cortex.
• The proposed shape-based loss function improved state-of-the-art models, increasing MIoU by an average of 0.0156 (nucleus) and 0.0078 (cortex), and reducing MED by 1.394 (nucleus) and 0.134 (cortex) pixels.

Conclusions:

• Transformed lens segmentation from a classification to a regression task by learning a level set function.
• Successfully embedded shape information into deep learning through novel loss function design.
• The proposed method offers enhanced efficiency and accuracy for segmenting objects with weak-contrast boundaries in AS-OCT images.