Cross-Modality Multi-Atlas Segmentation via Deep Registration and Label Fusion

Area of Science:

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

Background:

• Multi-atlas segmentation (MAS) is a key technique for medical image segmentation.
• Conventional MAS relies on same-modality atlases, which are often scarce in clinical practice.
• Existing MAS methods face computational challenges in registration and label fusion.

Purpose of the Study:

• To develop an efficient cross-modality MAS framework for segmenting medical images using atlases from different modalities.
• To address the limitations of atlas availability and computational burden in conventional MAS.
• To improve the accuracy and efficiency of medical image segmentation through novel deep learning approaches.

Main Methods:

• A novel cross-modality MAS framework utilizing deep neural networks for both registration and label fusion.
• Bi-directional registration network (BiRegNet) for efficient alignment of cross-modality images.
• Similarity estimation network (SimNet) for adaptive label fusion based on multi-scale similarity assessment.

Main Results:

• The proposed framework demonstrated effectiveness in cross-modality MAS tasks.
• BiRegNet achieved efficient and accurate registration between different imaging modalities.
• SimNet improved label fusion performance by learning multi-scale similarity information.
• Successful evaluation on left ventricle and liver segmentation tasks using MM-WHS and CHAOS datasets.

Conclusions:

• The developed deep learning-based cross-modality MAS framework is effective and computationally efficient.
• This approach overcomes the limitations of same-modality atlas availability in medical image segmentation.
• The framework shows significant potential for improving clinical medical image analysis and segmentation accuracy.