Quantitative error prediction of medical image registration using regression forests

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Accurate medical image registration is crucial for clinical adoption of imaging techniques.
• Quantitative prediction of registration error can enhance image registration quality.
• Lack of ground truth in medical images presents challenges for predicting registration error.

Purpose of the Study:

• To propose an automatic method for quantitative prediction of medical image registration error.
• To apply the method to chest CT scans for evaluating registration accuracy.
• To enable automatic quality control in large-scale medical image analysis.

Main Methods:

• A random regression forest model was developed for local registration error prediction.
• Features used included transformation model parameters and post-registration dissimilarity.
• The model was trained and tested on chest CT scan datasets (SPREAD, DIR-Lab-4DCT, DIR-Lab-COPDgene).

Main Results:

• Mean absolute errors of 1.07 ± 1.86 mm (SPREAD) and 1.76 ± 2.59 mm (inter-database) were achieved.
• Classification accuracy for correct, poor, and wrong registration was 90.7% (SPREAD) and 75.4% (inter-database).
• The method demonstrated robust performance across different datasets.

Conclusions:

• The proposed automatic method effectively predicts medical image registration error quantitatively.
• The approach shows promise for applications like automated quality control in medical imaging.
• Accurate registration error prediction facilitates the clinical implementation of advanced imaging techniques.