Pancreas segmentation with probabilistic map guided bi-directional recurrent UNet.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Anatomy

Background:

• Pancreas segmentation is crucial for diagnostics and treatment but challenging due to anatomical variations.
• Current 2D methods lose temporal information, while 3D methods are computationally expensive.
• Fully convolutional neural networks (FCNs) struggle with pancreas segmentation variability.

Purpose of the Study:

• To propose a novel Probabilistic-Map-guided Bi-directional Recurrent UNet (PBR-UNet) architecture.
• To address the limitations of existing 2D and 3D segmentation methods for the pancreas.
• To achieve accurate and computationally efficient pancreas segmentation.

Main Methods:

• Developed a PBR-UNet integrating intra-slice and inter-slice probabilistic maps.
• Utilized a local 3D hybrid regularization scheme with bi-directional recurrent optimization.
• Employed an initial estimation module for pixel-level probabilistic maps and a 2.5D UNet for information propagation.

Main Results:

• The PBR-UNet effectively fuses local 3D information using adjacent slice probabilistic maps.
• Bi-directional recurrent optimization enhances the utilization of local context.
• Achieved comparable segmentation accuracy to state-of-the-art methods with reduced computational cost on NIH Pancreas-CT and MSD datasets.

Conclusions:

• The PBR-UNet offers an efficient and effective solution for pancreas segmentation.
• This approach overcomes the limitations of traditional 2D and 3D segmentation techniques.
• The method demonstrates significant potential for clinical applications in pancreas diagnostics.