Uncertainty estimation and evaluation of deformation image registration based convolutional neural networks

Area of Science:

• Medical Imaging
• Radiotherapy
• Machine Learning

Background:

• Accurate deformable image registration (DIR) and uncertainty estimation are vital for safe clinical deployment.
• Current deep learning models for DIR face challenges in uncertainty evaluation and hyperparameter optimization.
• This research addresses the need for rapid DIR with reliable uncertainty prediction.

Purpose of the Study:

• To develop and evaluate a novel probabilistic multi-resolution deep learning model for fast DIR and uncertainty estimation.
• To introduce a new metric based on Kullback-Leibler divergence for assessing predicted displacement field distribution quality.
• To compare the proposed model against existing uncertainty-predicting DIR algorithms.

Main Methods:

• A probabilistic multi-resolution convolutional neural network model was developed to estimate a multivariate normal distributed dense displacement field (DDF).
• Kullback-Leibler divergence was used to evaluate the quality of the predicted DDF.
• The model was applied to register planning CT to cone beam CT for adaptive radiotherapy and compared with VoxelMorph, Monte Carlo dropout, and Monte Carlo B-spline methods.

Main Results:

• Hyperparameter tuning revealed a trade-off between uncertainty reliability and deformation accuracy.
• The proposed model demonstrated superior performance in contour propagation and uncertainty estimation (p < 0.05) compared to other methods at the optimal trade-off.
• Achieved an average Dice Similarity Coefficient of 0.89 and a KL-divergence of 0.15.

Conclusions:

• The developed model reliably predicts both deformable image registration and its uncertainty.
• This addresses key challenges in DIR uncertainty estimation and evaluation.
• The findings pave the way for the safe clinical deployment of deep learning-based DIR.