Adjusting the Ground Truth Annotations for Connectivity-Based Learning to Delineate.

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Accurate 3D structure delineation is crucial for deep learning models.
• Manual annotation in 3D is challenging due to visual interpretation difficulties and awkward interfaces.
• Existing methods struggle with inherent inaccuracies in large-scale 3D annotations.

Purpose of the Study:

• To develop a method that explicitly accounts for and corrects annotation inaccuracies in 3D structure delineation.
• To improve the performance of deep learning networks trained on potentially flawed 3D data.
• To enable joint training of networks and correction of annotation errors.

Main Methods:

• Proposed a novel approach treating annotations as active contour models.
• Implemented a method that allows annotations to deform while preserving topology.
• Developed a technique for jointly training deep networks and refining annotations.

Main Results:

• The proposed method successfully accounts for annotation inaccuracies.
• Joint training with deformable contour models improved network performance.
• The approach demonstrated enhanced accuracy in 3D structure delineation despite initial annotation errors.

Conclusions:

• Explicitly modeling annotation inaccuracies enhances deep learning performance for 3D delineation.
• Active contour models offer a robust way to refine annotations during training.
• This method provides a more reliable approach to training deep networks with real-world, imperfect 3D data.