Modality-agnostic self-supervised deep feature learning and fast instance optimisation for multimodal fusion in...

Area of Science:

• Medical Imaging
• Image Registration
• Deep Learning

Background:

• Accurate alignment of pre-operative MRI scans to intra-operative ultrasound is crucial for image-guided interventions.
• Existing learning-based methods struggle with both speed and robustness to large motion/misalignment.

Purpose of the Study:

• To develop a novel method for fast and robust 3D multimodal medical image registration.
• To decouple feature learning from transformation prediction for improved performance.

Main Methods:

• A convolutional neural network (CNN) extracts modality-agnostic features for fast inference.
• Sparsity-based pruning reduces model complexity and computation time.
• Instance optimization robustly estimates global transformation from local displacement probabilities.

Main Results:

• Achieved state-of-the-art accuracy with an average target registration error of 2.50 mm on the CuRIOUS dataset.
• The model has a small size (1.2 MByte) and runs in approximately 3 seconds for full 3D registration.
• Demonstrated robustness to large motion and misalignment.

Conclusions:

• The proposed method significantly improves accuracy and robustness in medical image registration.
• Fast self-supervised deep learning enables state-of-the-art performance on challenging registration tasks.