FIND: A Framework for Iterative to Non-Iterative Distillation for Lightweight Deformable Registration.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Deformable image registration is vital for medical image analysis but computationally intensive.
• Existing deep learning methods struggle with deployment on resource-limited devices.
• Current knowledge distillation techniques are ineffective for transferring complex deformation capabilities to lightweight networks.

Purpose of the Study:

• To propose a novel framework, FIND, for efficient knowledge distillation in deformable image registration.
• To enable lightweight networks to handle complex deformations effectively.
• To facilitate the deployment of advanced registration models on resource-constrained hardware.

Main Methods:

• Propose the Framework for Iterative to Non-iterative Distillation (FIND).
• Employ a dual-step distillation process: recurrent distillation to create a teacher assistant, followed by feature distillation to a lightweight network.
• Utilize a Non-Iterative Lightweight (NIL) network for rapid and effective registration.

Main Results:

• The NIL network, trained with FIND, achieves significantly faster performance (up to 60x on CPU, 89x on GPU) compared to existing deep learning methods.
• Demonstrates superior registration accuracy, with improvements up to 3.5 Dice score points.
• Successfully enables effective registration on resource-limited devices.

Conclusions:

• FIND provides an efficient method for transferring complex deformation handling from iterative to non-iterative lightweight networks.
• The resulting NIL network offers a practical solution for real-time medical image registration applications.
• This approach significantly enhances the accessibility and applicability of deep learning-based registration techniques.