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

2025-06-02

AI Summary

Clinical CT reconstruction under low-dose, sparse-view, or metal-implant conditions suffers from severe noise, prominent artifacts, and the impracticality of acquiring ground-truth paired training data. Method: This paper proposes a physics-informed generative reconstruction framework that operates without precisely paired training data. It innovatively integrates the Poisson Flow Generative Model (PFGM++) into the CT reconstruction pipeline, tightly coupling a differentiable physical projection operator with a jointly optimized Reconstruction Conditional Engine (FORCE), enabling end-to-end synergy between data-driven priors and physical fidelity constraints. Contribution/Results: The method avoids hallucination risks arising from data inconsistency in conventional deep learning approaches. It significantly outperforms existing unsupervised methods across multiple challenging scenarios, demonstrating superior noise suppression, structural fidelity, and cross-domain generalizability.

Technology Category

• Computer Vision: Low Level & Physics-based Vision
• Machine Learning: Deep Generative Models & Autoencoders
• Natural Language Processing: Generation

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Economics, Online Markets and Human Computation: Trust and reliance of crowd workers and data experts on GenAI
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation

Abstract

Computed tomography (CT) is a major medical imaging modality. Clinical CT scenarios, such as low-dose screening, sparse-view scanning, and metal implants, often lead to severe noise and artifacts in reconstructed images, requiring improved reconstruction techniques. The introduction of deep learning has significantly advanced CT image reconstruction. However, obtaining paired training data remains rather challenging due to patient motion and other constraints. Although deep learning methods can still perform well with approximately paired data, they inherently carry the risk of hallucination due to data inconsistencies and model instability. In this paper, we integrate the data fidelity with the state-of-the-art generative AI model, referred to as the Poisson flow generative model (PFGM) with a generalized version PFGM++, and propose a novel CT framework: Flow-Oriented Reconstruction Conditioning Engine (FORCE). In our experiments, the proposed method shows superior performance in various CT imaging tasks, outperforming existing unsupervised reconstruction approaches.

Problem

• Improving CT image reconstruction in low-dose, sparse-view, and metal implant scenarios
• Addressing hallucination risks from data inconsistencies in deep learning methods
• Proposing FORCE framework integrating PFGM++ for superior unsupervised reconstruction

Innovation

• Integrates data fidelity with PFGM++
• Proposes FORCE for CT reconstruction
• Outperforms unsupervised reconstruction methods