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

Performance of deep learning-based models for tumour tracking in magnetic resonance-guided radiotherapy hinges on high-quality labelled data. Medical images are prone to annotation errors, making data cleaning essential. We propose an automatic data cleaning tool based on the foundation model Segment Anything 2, which incorporates temporal information from cinematic magnetic resonance imaging to detect annotation errors and generate corrected contours, minimising manual effort involved in data cleaning. Expert validation by two radiation oncologists showed a preference for corrected contours over the original manual contours. Corrected contours (DSC 0.95 ± 0.01) surpassed interobserver variability (0.88 ± 0.02) on a dataset annotated by five observers.