Automated annotation error detection and correction for manual two-dimensional cinematic magnetic resonance imaging...

Area of Science:

• Medical imaging
• Artificial intelligence
• Radiotherapy

Background:

• Deep learning models for tumor tracking in MRI-guided radiotherapy require high-quality labeled data.
• Medical image annotations are susceptible to errors, necessitating robust data cleaning methods.

Purpose of the Study:

• To introduce an automated data cleaning tool for enhancing the quality of labeled data in medical imaging.
• To leverage foundation models and temporal information for accurate annotation error detection and correction.

Main Methods:

• Utilized the Segment Anything 2 foundation model for image segmentation.
• Incorporated temporal information from cinematic MRI to identify and correct annotation inconsistencies.
• Developed an automated workflow for data cleaning and contour generation.

Main Results:

• The automated tool successfully detected annotation errors and generated corrected contours.
• Expert radiation oncologists preferred the AI-corrected contours over original manual annotations.
• Corrected contours achieved a Dice Similarity Coefficient (DSC) of 0.95 ± 0.01, exceeding interobserver variability (0.88 ± 0.02).

Conclusions:

• The proposed automatic data cleaning tool significantly improves the accuracy of tumor contours in MRI-guided radiotherapy.
• This approach minimizes manual effort in data cleaning and enhances the reliability of deep learning models for radiotherapy applications.