Predicting dice similarity coefficient of deformably registered contours using Siamese neural network

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence

Background:

• Deformable image registration (DIR) is crucial for adaptive radiotherapy, enabling dose accumulation.
• Manual contour evaluation for DIR is time-consuming and prone to inter-observer variability.
• Accurate registration of organs-at-risk (OARs) contours is essential for treatment planning.

Purpose of the Study:

• To develop a deep learning model for predicting Dice Similarity Coefficients (DSC) of registered OAR contours.
• To automate the evaluation of DIR accuracy in prostate radiotherapy.
• To reduce the burden of manual contour assessment.

Main Methods:

• A Siamese neural network was trained on virtual CT (vCT) and cone-beam CT (CBCT) image pairs.
• The model predicted DSC, a metric for contour registration quality.
• Performance was evaluated using Root Mean Squared Error (RMSE) and classification accuracy.

Main Results:

• The model achieved low RMSE for rectum (0.070), prostate (0.079), and bladder (0.118).
• It demonstrated 92% accuracy in classifying rectal contour registration as good or poor (DSC < 0.6).
• A sensitivity of 0.97 indicates high accuracy in identifying poorly registered contours.

Conclusions:

• The proposed deep learning model accurately predicts DSC for registered OAR contours.
• This tool can reliably assess DIR quality and identify contours requiring physician review.
• The model facilitates efficient plan adaptation in radiotherapy.