Quality assurance for automatically generated contours with additional deep learning.

Area of Science:

• Medical Imaging AI
• Deep Learning for Segmentation Quality Assurance

Background:

• Automatic segmentation models require rigorous quality assurance (QA), especially in healthcare.
• Current tools for AI model QA are lacking for researchers.
• This work addresses the need for automated QA in medical image segmentation.

Purpose of the Study:

• To develop a deep learning model for estimating the quality of automatically generated contours.
• To provide AI researchers with tools for QA in high-stakes healthcare applications.

Main Methods:

• A 3D EfficientDet architecture with a regression head was trained to predict Dice similarity coefficient from image-contour pairs.
• A dataset of 60 prostate MRI images with ground truth and 80 automatic segmentations was used.
• Extensive data augmentation and fivefold cross-validation were employed to address dataset limitations.

Main Results:

• The model achieved a mean absolute error of 0.020 ± 0.026 in estimating Dice scores.
• A rank correlation of 0.42 was observed.
• The model correctly identified incorrect segmentations 99.6% of the time.

Conclusions:

• The trained deep learning model can enhance quality assurance for automatic segmentation tools.
• This enables intervention to prevent undesired segmentation behavior in clinical practice.