DINO-Reg: Efficient Multimodal Image Registration With Distilled Features.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Medical image registration is vital for aligning anatomical structures in 3D.
• Existing methods often require extensive fine-tuning or are computationally intensive.
• Vision foundation models offer powerful feature extraction capabilities.

Purpose of the Study:

• To introduce DINO-Reg, an adaptation-free registration method using the DINOv2 foundation model for 3D medical images.
• To develop DINO-Reg-Eco, a computationally efficient version via knowledge distillation.
• To evaluate the performance of these methods against existing registration techniques.

Main Methods:

• Leveraging the DINOv2 vision foundation model for feature extraction in deformable 3D medical image alignment.
• Proposing DINO-Reg-Eco, a knowledge-distilled model using a UNet-structured 3D CNN.
• Benchmarking DINO-Reg and DINO-Reg-Eco on diverse medical imaging datasets.

Main Results:

• DINO-Reg demonstrates state-of-the-art performance by generalizing DINOv2 to medical images without fine-tuning.
• DINO-Reg-Eco achieves a 99% reduction in encoding time, maintaining competitive performance.
• Both methods outperform current supervised and unsupervised registration approaches.

Conclusions:

• Foundation models, like DINOv2, can be effectively adapted for medical image registration.
• DINO-Reg and DINO-Reg-Eco offer efficient and high-performance solutions for 3D medical image alignment.
• These methods hold transformative potential for medical imaging applications, especially in resource-constrained environments.