Tomographic Foundation Model-FORCE: Flow-Oriented Reconstruction Conditioning Engine

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Reconstruction

Background:

• Computed tomography (CT) is crucial for medical diagnostics, but clinical challenges like low-dose scanning and metal artifacts degrade image quality.
• Deep learning significantly improves CT reconstruction, yet acquiring paired training data is difficult, leading to potential image hallucination.
• Existing unsupervised methods struggle with data inconsistencies and model instability in CT image reconstruction.

Purpose of the Study:

• To develop a novel CT image reconstruction framework that addresses the challenges of data scarcity and image artifacts.
• To leverage state-of-the-art generative AI, specifically Poisson flow generative models (PFGM/PFGM++), for improved CT reconstruction.
• To enhance the robustness and accuracy of CT image reconstruction in clinical settings.

Main Methods:

• Integration of data fidelity principles with the Poisson flow generative model (PFGM++) to create the FORCE framework.
• Development of a novel CT reconstruction framework named FORCE (Flow-Oriented Reconstruction Conditioning Engine).
• Experimental validation of FORCE across various CT imaging tasks, including low-dose and sparse-view scenarios.

Main Results:

• The proposed FORCE framework demonstrates superior performance in CT image reconstruction tasks.
• FORCE effectively mitigates noise and artifacts commonly encountered in clinical CT imaging.
• The method outperforms existing unsupervised CT reconstruction approaches in experimental evaluations.

Conclusions:

• FORCE offers a significant advancement in CT image reconstruction, particularly for challenging clinical scenarios.
• The framework successfully combines generative AI with data fidelity to produce high-quality reconstructed images.
• FORCE represents a promising unsupervised approach for improving CT imaging without the need for perfectly paired data.