Abstract
Multiphase contrast-enhanced computed tomography (CECT) scan is clinically significant to demonstrate the anatomy at different phases. In practice, such a multiphase CECT scan inherently takes longer time and deposits much more radiation dose into a patient body than a regular CT scan, and reduction of the radiation dose typically compromise the CECT image quality and its diagnostic value. With Joint Condition and Circle-Supervision, here we propose a novel Poisson Flow Generative Model (JCCS-PFGM) to promote the progressive low-dose reconstruction for multiphase CECT. JCCS-PFGM is characterized by the following three aspects: a progressive low-dose reconstruction scheme, a circle-supervision strategy, and a joint condition mechanism. Our extensive experiments are performed on a clinical dataset consisting of 11436 images. The results show that our JCCS-PFGM achieves promising PSNR up to 46.3dB, SSIM up to 98.5%, and MAE down to 9.67 HU averagely on phases I, II and III, in quantitative evaluations, as well as gains high-quality readable visualizations in qualitative assessments. All of these findings reveal our method a great potential to be adapted for clinical CECT scans at a much-reduced radiation dose.
Introduction
Multiphase Contrast-Enhanced Computed Tomography (CECT) is a cornerstone of modern radiology, used to visualize blood flow and organ function by scanning a patient multiple times (e.g., arterial, venous, and delayed phases) after injecting a contrast agent.
The catch? Each additional "phase" adds more ionizing radiation. A triple-phase scan can deposit three times the radiation of a standard CT, increasing long-term cancer risks. While we can lower the dose, doing so typically results in "noisy" images filled with artifacts that make them difficult for radiologists to read. Most existing AI denoising tools treat each phase in isolation, failing to exploit the fact that the patient's anatomy doesn't change between scans—only the contrast distribution does.
Methodology: The JCCS-PFGM Framework
The authors introduce JCCS-PFGM (Joint Condition and Circle-Supervision based Poisson Flow Generative Model). It moves away from standard denoising and toward a sophisticated generative approach based on three pillars:
- Progressive Reconstruction Scheme: Instead of an even dose distribution, the authors assign a higher dose to the first phase (30%) to establish a clear anatomical "anchor" and progressively lower doses to subsequent phases (15% and 5%).
- Poisson Flow Generative Model (PFGM): Unlike popular Diffusion Models that mimic gas moving through a room, PFGM is inspired by electrostatics. It treats image data points as electrical charges on a hyperplane, creating a "Poisson field." This approach is often more robust and computationally efficient for mapping noisy data back to clean images.
- Circle-Supervision Strategy: To prevent the model from "hallucinating" details, the authors implement a self-inspection loop. The model predicts a Poisson field, restores an image, perturbs it again, and checks if the new prediction matches the original. This "circle" ensures the model learns the actual underlying structure rather than just generic patterns.
- Joint Condition Mechanism: This uses attention-based "cross-fusion" to pull relevant anatomical details from previous phases into the current reconstruction, ensuring that the evolution of the contrast agent is the only thing changing, not the patient's organs.
Key Insights and Results
The study utilized a substantial clinical dataset of 38,496 CECT images from 247 patients. Low-dose versions were generated using a validated photon-counting model that simulates real-world physics like electronic noise and bowtie filters.
The JCCS-PFGM outperformed standard reconstruction (FBP) and state-of-the-art deep learning models (RED-CNN, DDPNet) across all metrics:
- Quantitative Accuracy: Achieved a Peak Signal-to-Noise Ratio (PSNR) of up to 46.3dB and a Structural Similarity Index (SSIM) of 98.5%.
- Error Reduction: The Mean Absolute Error (MAE) dropped to 9.67 HU (Hounsfield Units) on average across all phases.
- Ultra-Low Dose Performance: Even at the 5% dose level (Phase III), the model preserved tiny features like small blood vessels that were completely lost or blurred in competing methods.
Limitations and Future Work
While the simulation model is highly accurate, the framework still needs to be validated on hardware-native low-dose scans where "real-world" artifacts might differ slightly. Additionally, the model assumes the patient hasn't moved significantly between phases. In clinical practice, respiratory motion can occur, which might require integrating a motion-correction step into the joint condition mechanism. Future work could also explore how this model performs on different organ systems (e.g., cardiac vs. abdominal) where contrast dynamics vary significantly.
Summary updated