Contour interpolation by deep learning approach.

Area of Science:

• Medical Imaging
• Computational Anatomy
• Artificial Intelligence in Medicine

Background:

• Contour interpolation is crucial for expediting manual segmentation of anatomical structures in medical imaging.
• Conventional shape-based interpolation (SBI) methods often show suboptimal performance, particularly for organ borders and gastrointestinal structures.

Purpose of the Study:

• To develop and validate a generic deep learning solution for robust and accurate contour interpolation.
• To address the limitations of traditional methods in challenging segmentation scenarios.

Main Methods:

• A deep contour interpolation model was developed and trained on a large dataset (16,796 cases, 15 organs).
• The model utilized image patches and 2D contour masks from top and bottom slices to predict organ masks for intermediate slices.
• Performance was evaluated using Dice scores and distance-to-agreement (DTA) metrics.

Main Results:

• The deep learning model achieved superior performance compared to conventional SBI, with a Dice score of 0.89 and mean DTA of 2.11 mm across 3167 test cases.
• Significant improvements were observed for difficult cases, with Dice scores of 0.85 and DTA of 3.45 mm, outperforming SBI (Dice: 0.78, DTA: 5.67 mm).
• Statistical analysis confirmed significant performance enhancements (p < 0.001) for Dice scores and DTA in challenging cases.

Conclusions:

• A novel deep learning approach effectively enhances contour interpolation in medical images.
• This method offers a valuable tool for accelerating manual segmentation tasks and improving accuracy for anatomical structures.