Project Introduction: High-Fidelity 3D Photon-Counting CT Denoising via Advanced Generative AI

1. Background and Motivation

Photon-counting computed tomography (PCCT) represents the next major technological leap in medical imaging. By directly converting X-ray photons into electrical signals, PCCT offers unprecedented spatial resolution, reduced electronic noise, and intrinsic multi-energy capabilities. However, clinical implementation must always balance image quality with patient safety. To minimize radiation exposure, Low-Dose PCCT (LD-PCCT) scans are highly desirable in clinical workflows.

Unfortunately, reducing the radiation dose inevitably introduces severe quantum noise and streak artifacts, which can obscure critical anatomical structures and micro-lesions, thereby compromising diagnostic accuracy. Traditional denoising algorithms and early-generation deep learning methods (such as standard CNNs) often struggle with this challenge, frequently resulting in an "over-smoothed" appearance that destroys vital physiological textures and alters Hounsfield Unit (HU) statistics.

To address this, there is a pressing need for a paradigm shift in medical image restoration. Advanced generative AI models are capable of learning complex data distributions rather than simply minimizing pixel-wise errors, which offers a promising solution. By mapping the intricate relationship between noisy and clean image domains, generative frameworks can synthesize high-fidelity textures and reconstruct anatomically accurate volumes without strictly relying on perfectly paired data constraints.

2. Proposed Framework & Objectives

This collaborative research project, undertaken by the Pattern Recognition Lab at FAU, Siemens Healthineers, and Peking Union Medical College Hospital (PUMCH), aims to develop a robust, end-to-end 3D AI framework for PCCT image restoration. Leveraging a highly valuable clinical dataset of 400 3D PCCT volumes, the project focuses on constructing a latent-space 3D generative model.

By compressing the massive high-resolution 3D volumes into a compact, anatomy-aware latent space, the framework mitigates the extreme GPU memory constraints typically associated with 3D medical data. Within this latent space, the generative model will learn the underlying distribution dynamics to organically guide low-dose inputs toward the high-quality, normal-dose target distribution, effectively restoring anatomical integrity while preserving accurate clinical metrics.

3. Research Questions

To guide the development and evaluation of this framework, this project addresses the following core research questions (RQs):

• RQ1: How can advanced 3D generative models be effectively adapted to reconstruct high-fidelity PCCT volumes under ultra-low radiation conditions without introducing artificial hallucinations?

• RQ2: How can latent-space representations be optimized to handle the extreme computational demands of high-resolution 3D medical data while preserving continuous axial structure and physiological textures?

• RQ3: To what extent do data-driven, distribution-matching generative paradigms outperform traditional supervised deep learning baselines in rigorous clinical evaluation metrics (e.g., LPIPS, HU deviation, and structural continuity)?

4. References

1. Willemink, M. J., et al. (2018). "Photon-counting CT: Technical Principles and Clinical Prospects." Radiology, 289(2), 293-312. (Highlights the hardware and clinical background of PCCT).

2. Kazerouni, A., et al. (2023). "Diffusion Models in Medical Imaging: A Comprehensive Survey." Medical Image Analysis, 88, 102846. (Provides context on the shift toward generative AI for medical image restoration).

3. Yang, Q., et al. (2018). "Low-Dose CT Image Denoising Using a Generative Adversarial Network with Wasserstein Distance and Perceptual Loss." IEEE Transactions on Medical Imaging, 37(6), 1348-1357. (Serves as a foundation for the limitations of earlier supervised and adversarial denoising methods).