Machine Learning-Based Quality Assurance for Automatic Segmentation of Head-and-Neck Organs-at-Risk in Radiotherapy.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiotherapy Oncology

Background:

• Deep learning, particularly convolutional neural networks (CNNs), is increasingly used for automatic segmentation in oncology.
• Evaluating the quality of CNN-generated contours is time-consuming for oncologists and often requires a gold standard.
• Existing evaluation metrics like Dice Similarity Coefficient (DSC) necessitate a ground truth for comparison.

Purpose of the Study:

• To develop an automatic quality assurance (QA) method for analyzing segmentation contour quality.
• To enable contour quality assessment without relying on a gold standard.
• To identify and locate poor-quality contours for further review.

Main Methods:

• A CNN segmentation network generated contours for 18 head-and-neck organs-at-risk across 196 individuals.
• Inner/outer shells were created from contours, and 38 radiomics features were extracted.
• Machine learning models used radiomics features and DSCs to classify slice quality; an anisotropic method identified poor contour locations.

Main Results:

• Isotropic experiments showed predicted values closely aligned with ground truth labels.
• The anisotropic method successfully provided location information for poor contours by analyzing peak-to-peak and area-to-area ratios.
• The QA method demonstrated qualitative prediction of segmentation quality.

Conclusions:

• The proposed automatic QA method effectively assesses segmentation quality without a gold standard.
• The method provides valuable location information for identifying and rectifying suboptimal contours.
• This approach can reduce oncologists' workload in evaluating segmentation accuracy.