Improving performance of medical image alignment through super-resolution.

Area of Science:

• Medical Imaging
• Computer Vision
• Radiology

Background:

• Medical image alignment is vital for monitoring patient health over time.
• Low-dose Cone-beam CT (CBCT) imaging quality and patient specifics can hinder alignment accuracy.

Purpose of the Study:

• To develop an unsupervised medical image alignment method.
• To enhance alignment accuracy by incorporating super-resolution preprocessing.
• To improve the tracking of patient conditions.

Main Methods:

• Proposed an unsupervised alignment method with super-resolution preprocessing.
• Developed a novel loss function combining Pearson correlation coefficient and regional mutual information.
• Validated the method on private clinical, public clinical, and public datasets.

Main Results:

• Higher resolution data demonstrably improved alignment results across all experiments.
• The new loss function outperformed common loss functions in alignment accuracy.
• Super-resolution preprocessing combined with the new loss function boosted alignment by up to 9.58% across different models.

Conclusions:

• The proposed unsupervised alignment method with super-resolution preprocessing effectively enhances medical image alignment accuracy.
• This advancement is crucial for reliable tracking of patient conditions.
• The method shows broad applicability across different alignment models.