UNet-based multi-organ segmentation in photon counting CT using virtual monoenergetic images

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Multi-organ segmentation is crucial for disease diagnosis, treatment planning, and radiotherapy.
• Photon counting detector-based CT (PCCT) offers spectral information that can potentially enhance segmentation performance.

Purpose of the Study:

• To propose and evaluate a UNet-based multi-organ segmentation method for PCCT utilizing virtual monoenergetic images (VMIs).
• To effectively leverage spectral information from PCCT for improved segmentation accuracy.

Main Methods:

• A multi-step process involving noise reduction, material decomposition, VMI generation, and deep learning-based segmentation (3D UNet, Swin UNETR).
• VMIs were synthesized across various x-ray energies using basis images.
• Segmentation performance was evaluated using Dice Similarity Coefficients (DSC) and 3D visualization on abdominal phantoms.

Main Results:

• The proposed VMI-based method demonstrated improved training stability compared to conventional bin-wise image segmentation.
• Average DSC for liver, pancreas, and spleen segmentation slightly increased from 0.933 to 0.95, with a reduced standard deviation from 0.066 to 0.047.
• Improvements were noted particularly in low-dose scenarios and with fewer energy bins.

Conclusions:

• Virtual monoenergetic images (VMIs) enhance training stability for multi-organ segmentation in PCCT.
• The VMI approach is particularly beneficial when working with a limited number of energy bins, offering a more robust segmentation solution.