Learning non-linear patch embeddings with neural networks for label fusion

Area of Science:

• Neuroimaging
• Computational Anatomy
• Machine Learning

Background:

• Multi-atlas strategies enhance brain structural segmentation by accommodating anatomical variability.
• Patch-based label fusion (PBLF) relies on local image similarity for weighted voting.

Purpose of the Study:

• To improve discriminative capabilities of similarity measurements in PBLF using neural network-computed patch embeddings.
• To evaluate the performance of different complexity embedding models.

Main Methods:

• A novel framework for computing patch embeddings via neural networks was developed.
• Embeddings included simple scaling, affine, and non-linear transformations.
• Methods were compared against state-of-the-art alternatives using hippocampus segmentation datasets.

Main Results:

• Even the simplest patch embedding versions outperformed standard PBLF.
• More complex transformation models yielded superior results.
• A considerable increase in average Dice score was observed compared to standard PBLF.

Conclusions:

• Discriminative learning through neural network patch embeddings significantly enhances PBLF for brain segmentation.
• The proposed framework offers a robust improvement over existing methods, particularly for complex anatomical structures like the hippocampus.