Prior guided deep difference meta-learner for fast adaptation to stylized segmentation.

Area of Science:

• Medical imaging and radiotherapy
• Artificial intelligence in healthcare
• Computational anatomy

Background:

• Radiotherapy treatment planning necessitates accurate segmentation of anatomical structures.
• Deep learning auto-segmentation models often fail to match diverse clinical segmentation styles.
• Adapting pre-trained models to new institutional styles is resource-intensive.

Purpose of the Study:

• To develop a method for adapting pre-trained auto-segmentation models to new, unseen clinician segmentation styles.
• To enable precise segmentation that aligns with local preferences without extensive retraining.
• To improve the efficiency and accuracy of radiotherapy contouring.

Main Methods:

• Proposed a Prior-guided deep difference meta-learner (DDL) to learn and adapt segmentation style differences.
• Utilized a dataset of 440 patients for development and 30 for testing, including prostate CTV, parotid, and rectum contours.
• Evaluated performance using Dice Similarity Coefficient (DSC) and Hausdorff distance, comparing with transfer learning.

Main Results:

• The DDL model adapted to new styles with minimal prior patient data (as few as 3 patients).
• Significant improvements in average DSC were observed across various structures (e.g., CTV, parotid, rectum).
• The model achieved high accuracy, outperforming transfer learning in adapting to unseen styles.

Conclusions:

• The Prior-guided DDL offers a fast and effortless solution for adapting segmentation models to new styles.
• Improved segmentation accuracy can reduce manual contour editing time, enhancing clinical workflow efficiency.
• This approach facilitates the deployment of auto-segmentation tools in diverse clinical settings.