Deep learning for contour quality assurance for RTOG 0933: In-silico evaluation

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Radiation Oncology

Background:

• Accurate hippocampal segmentation is crucial for radiotherapy planning to minimize dose to organs at risk.
• Current multi-institutional trials face challenges in maintaining consistent contouring quality across different centers.
• Deep learning (DL) models offer a potential solution for automating and standardizing segmentation tasks.

Purpose of the Study:

• To validate a CT-based deep learning (DL) model for hippocampal segmentation trained on single-institution data.
• To assess the DL model's utility for multi-institutional contour quality assurance (QA).

Main Methods:

• A DL model was trained on institutional observer (IO) contours from brain MRIs.
• The model was evaluated on the RTOG 0933 dataset, comparing DL contours with treating physician (TP) and IO contours using Dice and Hausdorff distance (HD).
• The DL model's ability to detect planning discrepancies was quantified using HD > 7 mm and Dmax > 17 Gy criteria.

Main Results:

• The DL model demonstrated superior agreement with IO contours (Dice 74%/73%) compared to TP contours (Dice 62%/65%).
• Thirty percent of contours and 53% of dose plans failed QA.
• The DL model achieved high AUC values (0.80/0.79 for contours, 0.91 for dose) in identifying QA failures.

Conclusions:

• A single-institution trained DL model is feasible for multi-institutional contour QA in hippocampal segmentation.
• The DL model shows promise in improving consistency and identifying discrepancies in multi-institutional radiotherapy trials.
• This approach can enhance the reliability of data and outcomes in large-scale clinical studies.