Multi-Atlas Image Soft Segmentation via Computation of the Expected Label Value

Area of Science:

• Medical Imaging
• Computer Vision
• Computational Anatomy

Background:

• Atlas-based segmentation is crucial in medical imaging.
• Deformable registration of atlases is computationally intensive and prone to local optima.
• Current methods often rely on a single optimal transformation, potentially missing valuable information.

Purpose of the Study:

• To propose a novel method for atlas-based medical image segmentation.
• To avoid the computational costs and local optima issues associated with deformable registration.
• To introduce the Expected Label Value (ELV) computation for improved segmentation accuracy.

Main Methods:

• Instead of performing deformable registration, we consider probabilities of all possible atlas-to-image transformations.
• We compute the Expected Label Value (ELV) by averaging over these probabilities.
• The ELV approach bypasses the need for explicit deformable registration.

Main Results:

• The proposed ELV computation method was evaluated on brain, liver, and pancreas segmentation.
• The approach was tested using both magnetic resonance (MR) and computed tomography (CT) imaging datasets.
• Results demonstrate the feasibility and potential benefits of the ELV method in diverse segmentation scenarios.

Conclusions:

• The Expected Label Value (ELV) offers a computationally efficient alternative to traditional deformable registration in atlas-based segmentation.
• This method mitigates the risk of local optima by considering a distribution of transformations.
• The ELV approach shows promise for improving the robustness and efficiency of medical image segmentation across different modalities and organs.