Automatic contouring QA method using a deep learning-based autocontouring system.

Area of Science:

• Medical Physics
• Radiotherapy
• Image Analysis

Background:

• Accurate contouring of target volumes and organs at risk is crucial for effective radiotherapy planning.
• Automated contouring systems offer potential efficiency gains but require robust verification methods.
• Evaluating the clinical acceptability of auto-generated contours is essential to ensure patient safety and treatment efficacy.

Purpose of the Study:

• To identify the most accurate similarity metric for verifying automatically generated contours using an independent system.
• To assess the performance of various similarity metrics in distinguishing clinically acceptable from unacceptable auto-contours.
• To establish reliable quantitative measures for quality assurance in automated radiotherapy contouring.

Main Methods:

• Utilized reference and verification autocontouring systems to generate paired contours for six pelvic structures across multiple CT datasets.
• Manually generated clinically acceptable and unacceptable contours for comparison against auto-generated ones.
• Calculated eleven similarity metrics, including Dice Similarity Coefficients (DSC) and Hausdorff distances, and employed Support Vector Machines (SVM) for classification.

Main Results:

• Surface DSC with 1-3mm thickness achieved high detection accuracies (0.89-0.97) for all evaluated structures.
• The linear kernel demonstrated consistent performance when combining metrics for SVM input.
• No combination of metrics outperformed the standalone surface DSC in terms of model accuracy.

Conclusions:

• Surface DSC with a 1-3mm thickness is the most accurate metric for differentiating clinically acceptable from unacceptable auto-contours in cervical cancer patients.
• This metric achieved over 0.9 accuracy for targets and critical structures, supporting its use in radiotherapy quality assurance.
• The findings provide a validated method for independent verification of automated contouring in clinical practice.