Spectral virtual non-contrast imaging assisted by artificial intelligence segmentation.

Area of Science:

• Medical Imaging
• Computational Imaging
• Radiotherapy Physics

Background:

• Quantitative virtual non-contrast (VNC) methods aim to derive radiotherapy parameters from contrast-enhanced spectral CT, avoiding extra scans.
• The challenge lies in the ill-posed nature of tissue characterization with limited spectral data, necessitating advanced techniques.

Purpose of the Study:

• To adapt a Bayesian VNC method for arbitrary energy channels in spectral photon-counting CT (PCCT).
• To integrate prior anatomical knowledge from AI-based multi-organ segmentation into the VNC method.
• To generalize the method for improved quantitative parameter estimation.

Main Methods:

• Reformulated a Bayesian VNC method for multiple energies and integrated it with AI segmentation (TotalSegmentator).
• Applied the method to simulated contrast-enhanced dual-energy CT (DECT) and PCCT datasets.
• Estimated radiotherapy parameters like electron density and proton stopping power ratio (SPR), comparing them to ground truth.

Main Results:

• AI-based segmentation significantly improved parameter estimation accuracy for both DECT and PCCT, especially PCCT.
• High spectral resolution combined with anatomical priors reduced Root Mean Square errors in SPR and electron density.
• Segmentation-assisted PCCT demonstrated superior performance in water-equivalent path length (WEPL) error compared to other methods.

Conclusions:

• A flexible, AI-assisted Bayesian framework was developed for quantitative analysis of contrast-enhanced spectral CT.
• Integration of AI segmentation and generalization to PCCT improved tissue characterization.
• The findings highlight AI's potential for quantitative insights beyond DECT, warranting further clinical validation.