Encoding atlases by randomized classification forests for efficient multi-atlas label propagation.

Area of Science:

• Medical Image Analysis
• Computational Anatomy
• Machine Learning

Background:

• Multi-atlas label propagation (MALP) is crucial for medical image segmentation.
• Current MALP methods often require computationally expensive non-linear registrations and complex fusion schemes.
• This high computational cost limits scalability and efficient experimentation.

Purpose of the Study:

• To develop a computationally efficient MALP method.
• To reduce the running time and improve the scalability of atlas-based segmentation.
• To maintain high accuracy comparable to existing state-of-the-art methods.

Main Methods:

• Encoding individual atlases using randomized classification forests to create an Atlas Forest (AF).
• Utilizing a classifier-based encoding approach, differing from traditional patch-based or direct value pair methods.
• Performing a single registration per target image and simple averaging for fusion of probabilistic label estimates.

Main Results:

• Achieved accuracy within the range of state-of-the-art methods across four diverse databases.
• Demonstrated a significantly lower running time compared to conventional MALP approaches.
• Showcased efficient experimentation and straightforward incorporation of new scans without retraining.

Conclusions:

• The proposed Atlas Forest (AF) method offers a computationally efficient alternative for MALP.
• This approach balances accuracy with reduced computational demands, enhancing scalability.
• AF facilitates faster experimentation and easier integration of new data in atlas-based image analysis.