Deep learning-based automatic contour quality assurance for auto-segmented abdominal MR-Linac contours.

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Deep-learning auto-segmentation (DLAS) aims to simplify contouring in clinical settings but faces challenges in abdominal MRI for MR-guided online adaptive radiotherapy (MRgOART).
• Automated contour quality assurance (ACQA) integrated with automatic contour correction (ACC) can optimize ACC performance by focusing on inaccurate contours.
• ACQA aids in selecting contours from various DLAS tools or propagating contours from previous sessions.

Purpose of the Study:

• To present the performance of novel 3D deep learning-based ACQA models for evaluating DLAS contours in abdominal MRI during MRgOART.
• To assess the accuracy of these models in identifying acceptable versus edit-required contours and detecting false positives/negatives.

Main Methods:

• A 3D convolutional neural network (CNN) ACQA model was trained on pancreas and duodenum contours from abdominal MRIs (1.5 T MR-Linac).
• The training dataset included 103 abdominal MR image datasets with DL contours and quality ratings.
• An in-house tool classified contours as acceptable or edit-required, and model performance was evaluated on 34 independent datasets using confusion matrices.

Main Results:

• The ACQA model achieved high accuracy in predicting contour quality: 72.2% for acceptable and 83.6% for edit-required pancreas contours; 71.2% for acceptable and 89.6% for edit-required duodenum contours.
• The model demonstrated strong performance in identifying false positive and false negative DLAS contours: 93.75% and 99.7% for pancreas, and 95% and 98.9% for duodenum, respectively.

Conclusions:

• Developed 3D ACQA models effectively evaluate the quality of DLAS pancreas and duodenum contours on abdominal MRI.
• These models can be integrated into clinical workflows to streamline contour evaluation for MRgOART in abdominal malignancies.
• The ACQA models contribute to efficient and consistent contour assessment, supporting the advancement of adaptive radiotherapy.