Towards Effcient Label Fusion by Pre-Alignment of Training Data

Area of Science:

• Medical image analysis
• Computational anatomy
• Biomedical imaging

Background:

• Label fusion is a multi-atlas segmentation technique offering superior accuracy compared to single-atlas methods.
• However, label fusion incurs high computational costs due to numerous pairwise image registrations.

Purpose of the Study:

• To develop a computationally efficient label fusion method.
• To maintain segmentation accuracy while reducing computational burden.

Main Methods:

• A modified label fusion approach was developed using diffeomorphic groupwise registration to pre-register training images.
• The novel image undergoes a single registration to the template, with subsequent transformations computed by concatenating existing ones.

Main Results:

• The modified method achieved significantly improved computational efficiency.
• Segmentation results were statistically indistinguishable from the standard label fusion method.
• Cardiac MR data experiments validated the approach.

Conclusions:

• The nonparametric inference algorithm is the primary driver of label fusion's benefits, not the multiple pairwise registrations.
• This modified approach offers a practical and efficient alternative for medical image segmentation.