X-WIN: Building Chest Radiograph World Model via Predictive Sensing

Computed Tomography (CT) is essential for clinical diagnostics, but the associated ionizing radiation poses health risks. Low-dose CT (LDCT) reduces exposure but compromises image quality. Researchers have developed an artificial intelligence-based framework to enhance LDCT images using a real-world paired lung dataset called Patient-uLDCT, which features ultra-low-dose images at only 2% of the normal radiation dose.

Identifiable causal representation learning aims to reveal the causal relationships in data generation processes. In medical imaging, this can enhance the generalizability and robustness of features used for disease classification. This study introduces a method for grouping observations to learn identifiable representations for chest X-ray classification, demonstrating improved performance across various classification tasks by enforcing invariance to factors such as race and sex.

The Poisson-Conditioned Generative Model (PoCGM) has been proposed to enhance sparse-view computed tomography (CT) reconstruction. This model aims to mitigate the challenges of aliasing artifacts and loss of structural details that arise from reducing projection views, which is crucial for minimizing radiation exposure and improving temporal resolution. By reformulating the Poisson Flow Generative Model (PFGM++) into a conditional framework, PoCGM integrates sparse-view data during training and sampling, improving the quality of reconstructed images.

The paper introduces Knowledge-Informed Dynamic Optimal Transport (KIDOT), a framework aimed at improving medical image reconstruction from measurement data. Traditional deep learning methods often rely on paired measurements and high-quality images, leading to performance issues when applied to real prospective data due to the retrospective-to-prospective gap. KIDOT addresses this by conceptualizing reconstruction as a dynamic transport path, learning from unpaired data and ensuring consistency with imaging physics.