Uncertain prediction of deformable image registration on lung CT using multi-category features and supervised...

Area of Science:

• Medical imaging
• Computational anatomy
• Machine learning

Background:

• Assessing deformable registration uncertainty is crucial for clinical applications but traditionally manual and time-consuming.
• Existing methods lack automated and reliable uncertainty prediction, impacting the safety and reliability of registration outcomes.

Purpose of the Study:

• To propose and validate a novel automatic method for predicting deformable registration uncertainty.
• To improve the efficiency and accuracy of uncertainty assessment in medical image registration.

Main Methods:

• A supervised learning approach using a random forest regressor trained on multi-category features.
• Features include deformation field statistics, physiological realism, and image similarity.
• Spatial adaptive random perturbations were employed to enhance feature discriminability.

Main Results:

• The proposed method accurately predicts local registration uncertainty.
• Quantitative experiments on thoracic CT datasets demonstrated superior performance compared to baseline methods.
• The approach showed effectiveness for both iterative optimization-based and deep learning-based registration.

Conclusions:

• The developed automatic method offers a reliable and efficient solution for deformable registration uncertainty prediction.
• This technique has significant potential to enhance the accuracy and clinical utility of medical image registration.