A multi-modal vision-language pipeline strategy for contour quality assurance and adaptive optimization

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence in Medicine

Background:

• Accurate delineation of organs-at-risk (OARs) is crucial for effective radiotherapy planning.
• Manual review and correction of deep learning-generated OAR segmentations are time-consuming and operator-dependent.

Purpose of the Study:

• To propose and evaluate a novel contour quality assurance and adaptive optimization (CQA-AO) strategy.
• To automate the identification and correction of inaccurate auto-segmentations for OARs in radiotherapy.

Main Methods:

• Developed a CQA-AO strategy with three components: contour QA, unacceptable contour category analysis, and adaptive optimization correction.
• Utilized a dataset of 586 CT images and labels from nine institutions, focusing on brainstem, parotid, and mandible OARs.
• Integrated vision-language representations and convex optimization algorithms for adaptive contour correction.

Main Results:

• The CQA-AO strategy achieved high sensitivity, accuracy, and precision in contour QA tasks for all OARs.
• The strategy demonstrated robust performance in identifying unacceptable contour categories and their locations.
• Adaptive optimization significantly improved Dice Similarity Coefficient (DSC) values by up to 25.9% and reduced Hausdorff distance by up to 81.6%.

Conclusions:

• The proposed CQA-AO strategy offers superior performance compared to conventional methods for OAR contouring.
• This automated approach enhances the efficiency of radiotherapy contouring and reviewing workflows.
• The CQA-AO strategy shows potential for clinical implementation to improve radiotherapy planning accuracy.