This innovation introduces a novel score-matching optimization model derived through deep learning. By integrating a learned score function into the Maximum A Posteriori (MAP) estimation framework, we have developed an iterative image reconstruction method that significantly enhances diagnostic clarity. Validated on both public medical CT datasets and clinical raw data, our method demonstrates superior denoising and deblurring performance. It consistently outperforms state-of-the-art approaches with remarkable generalizability and stability, offering immense potential for broader applications in medical image restoration and generation.
While Dual-Energy CT (DECT) offers valuable material-selective data, it comes with increased hardware complexity and radiation exposure. Our invention overcomes these limitations by generating Virtual Monoenergetic (VM) images directly from standard single-spectrum CT scans. Utilizing an improved residual neural network (IResNet), our model maps single-spectrum input to VM images at specified energy levels with high precision—achieving a relative error of less than 0.2% compared to true DECT-derived images. This technology enables accurate multi-material decomposition without the need for specialized DECT hardware, offering a safer, cost-effective, and low-dose alternative for clinical diagnostics.
The integration of AI into medical imaging is catalyzing a paradigm shift in healthcare, redefining diagnostic accuracy and clinical workflows. Our strategic roadmap focuses on next-generation models capable of identifying subtle pathological patterns beyond human perception. By improving sensitivity and specificity, these future tools will enable the ultra-early detection of cancers, neurological disorders, and cardiovascular diseases. We are committed to developing technologies that support interventions at the most treatable stages, ultimately driving better patient outcomes.