Fast Multi-Organ Fine Segmentation in CT Images with Hierarchical Sparse Sampling and Residual Transformer.

Area of Science:

• Medical Imaging Analysis
• Deep Learning in Radiology
• Computational Anatomy

Background:

• Accurate multi-organ segmentation in 3D medical images is crucial for clinical automation.
• Voxel-by-voxel deep learning methods are computationally intensive, limiting speed.
• Existing fast classifiers face a trade-off between speed and segmentation accuracy.

Purpose of the Study:

• To develop a novel, fast multi-organ segmentation framework for 3D medical images.
• To address the computational limitations of current deep learning segmentation techniques.
• To improve both speed and accuracy in organ segmentation for clinical applications.

Main Methods:

• Proposed a framework combining hierarchical sparse sampling with a Residual Transformer network.
• Hierarchical sparse sampling reduces computation by analyzing multiple resolution levels.
• The Residual Transformer efficiently extracts and combines multi-level information from sparse data.

Main Results:

• Achieved improved qualitative and quantitative segmentation performance compared to existing fast classifiers.
• Demonstrated significantly reduced computation time, achieving segmentation in approximately 2.24 seconds on CPU.
• Validated on an internal dataset (10,253 CT images) and the public TotalSegmentator dataset.

Conclusions:

• The proposed framework offers a significant advancement in fast and accurate multi-organ segmentation.
• The method overcomes the speed-accuracy trade-off, enabling efficient analysis of 3D medical volumes.
• This approach shows strong potential for real-time fine organ segmentation in clinical settings, aiding tasks like registration and detection.