High-resolution 3D abdominal segmentation with random patch network fusion.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• 3D abdominal organ segmentation on high-resolution CT scans presents memory limitations for GPUs and complex 3D fully convolutional networks.
• Existing strategies like lower-resolution/wider field-of-view and higher-resolution/limited field-of-view yield varied results.

Purpose of the Study:

• To propose a novel, memory-conservative patch-based network for accurate 3D abdominal organ segmentation on high-resolution CT.
• To improve upon existing methods by employing random spatial initialization and statistical fusion on overlapping regions of interest (ROIs).

Main Methods:

• A patch-based network utilizing random spatial initialization and statistical fusion on overlapping ROIs is proposed.
• The method involves processing low-resolution sections for complete spatial information preservation, followed by interpolation and random patch sampling.
• High-dimensional probability maps are generated by integrating patches across fields of view.

Main Results:

• The proposed method achieved a mean Dice Similarity Coefficient (DSC) score of 0.856, significantly improving multi-organ segmentation from a baseline of 0.799 (p < 0.01).
• Performance was evaluated across three datasets with 260 subjects, demonstrating robustness with varying manual labels.
• The approach outperformed other state-of-the-art methods in abdominal organ segmentation accuracy.

Conclusions:

• The developed approach offers a memory-efficient framework for 3D segmentation of high-resolution CT data.
• It is compatible with various base network architectures without significantly increasing inference complexity.
• This method provides a viable solution for accurate and efficient abdominal organ segmentation in medical imaging.