Deep learning-based affine medical image registration for multimodal minimal-invasive image-guided interventions - A...

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Computational imaging

Background:

• Multimodal image registration integrates data from diverse imaging modalities, crucial for medical analysis.
• Existing neural network approaches for medical image registration lack standardized comparison due to varied datasets.
• Affine registration is a key technique for aligning medical images.

Purpose of the Study:

• To implement and evaluate 20 neural networks for affine multimodal medical image registration.
• To assess the performance and generalizability of these networks on synthetic and real patient datasets.
• To compare network performance against a custom CNN benchmark and SimpleElastix baseline.

Main Methods:

• Implementation of 20 distinct neural networks for affine medical image registration.
• Semi-supervised training on a synthetic 3D CT/MR liver dataset.
• Evaluation on both synthetic and unseen real patient datasets, followed by fine-tuning on patient data.

Main Results:

• Six networks significantly improved Dice coefficients on the synthetic dataset (p < 0.05).
• Nine networks demonstrated significant improvement on the patient dataset, indicating generalization capabilities.
• Performance was benchmarked against a developed CNN and SimpleElastix.

Conclusions:

• Several neural networks show promise for generalizable affine multimodal medical image registration.
• Current methods often lack broad applicability across different datasets and clinical scenarios.
• Further research is essential to develop robust and widely applicable medical image registration techniques.