Deformable image registration with deep network priors: a study on longitudinal PET images

Area of Science:

• Medical Imaging
• Radiology
• Computational Biology

Background:

• Longitudinal registration of PET imaging is crucial for monitoring metastatic breast cancer.
• Deep learning (DL) has shown limited success in improving image registration performance compared to conventional methods.

Purpose of the Study:

• To introduce a novel deep learning approach, MIRRBA, for longitudinal PET image registration.
• To bridge the performance gap between conventional and DL-based registration methods.

Main Methods:

• MIRRBA is a subject-specific deformable registration method using a deep pyramidal architecture.
• It optimizes network parameters using only the pair of images to be registered, eliminating the need for a learning database.
• Evaluated on 110 whole-body PET scans of metastatic breast cancer patients, compared against ANTs, Elastix, LapIRN, and Voxelmorph.

Main Results:

• MIRRBA significantly improved Dice scores for organs (6-5%) and lesions (52-65%) compared to DL methods (Voxelmorph, LapIRN).
• MIRRBA demonstrated comparable performance to conventional iterative methods (ANTs, Elastix).
• The method's ability to address shrinking lesions was also assessed.

Conclusions:

• MIRRBA offers a promising approach for accurate longitudinal PET image registration in metastatic breast cancer.
• The study highlights the regularizing power of deep architectures in DL-based image registration.
• MIRRBA provides new insights into the role of architecture in DL registration methods.