JS-RegNeXt: A ConvNeXt-based few-shot JSR framework with correlation awareness and multi-scale prediction consistency

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Label-constrained (LC) medical image registration struggles with insufficient labels, leading to overfitting.
• Joint segmentation and registration (JSR) methods show promise but lack global correlation awareness, impacting performance in low-contrast anatomies.
• Robust medical image registration is crucial for accurate diagnosis and treatment planning.

Purpose of the Study:

• To propose a novel JS-RegNeXt framework for few-shot label-constrained medical image registration.
• To enhance global semantic perception and correlation awareness for improved registration robustness.
• To mitigate segmentation uncertainty and improve performance in low-contrast regions.

Main Methods:

• Developed a JS-RegNeXt framework with integrated segmentation and registration modules.
• Designed a SegNet with multi-scale prediction consistency for robust semantic perception.
• Proposed a RegNeXt incorporating ConvNeXt's large receptive field for enhanced global perception and correlation awareness.

Main Results:

• JS-RegNeXt demonstrated improved performance in both segmentation and registration tasks on cardiac CT and brain MRI datasets.
• The framework showed enhanced robustness in low-contrast regions compared to state-of-the-art methods.
• Achieved more accurate and reliable medical image registration, particularly in few-shot scenarios.

Conclusions:

• The JS-RegNeXt framework offers a robust solution for few-shot label-constrained medical image registration.
• The integration of global semantic perception and correlation awareness significantly improves registration accuracy.
• JS-RegNeXt shows substantial potential for clinical applications in medical imaging.