Scholarlink

This study employs deep learning algorithms to minimize ionizing radiation exposure during radiotherapy planning while preserving essential diagnostic information. We utilized CycleGAN+U-Net and SynDiff models to generate synthetic computed tomography (sCT) images from T1-weighted MRI scans of 37 patients exhibiting a spectrum of neurological conditions, ranging from mild cognitive impairment to severe Alzheimer’s disease. Model development and evaluation employed a rigorous k-fold cross-validation strategy. The dataset was partitioned, reserving 4 subjects as an independent test set. The remaining 33 subjects were utilized in a five-fold cross-validation process, systematically rotating subsets for training and validation to optimize model generalizability and performance assessment. From the comparison of the performances of the two applied models, SynDiff obtained higher segmentation accuracy values of ME (1066.92), MAE (1066.92), MSE (1474882.42), RMSE (1212.81), PSNR (34.67) and SSIM (0.001) than Cycle Gan + Unet and can be used to generate accurate sCT from brain MRI images. A key advantage of this AI-based sCT generation is the significant reduction in patient ionizing radiation exposure. This approach facilitates the preferential use of MRI, leveraging its superior soft-tissue contrast to provide essential information for detecting, monitoring, and diagnosing intracranial pathologies

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Unsupervised MR to CT Synthesis: A Comparative Analysis of Explicit Structural Constrained Adversarial Learning and Adversarial Diffusion Models

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Luca Cruciata,
  • Muhammad Ali,
  • Salvatore Contino,
  • Viviana Benfante,
  • Roberto Pirrone,
  • Albert Comelli

摘要

This study employs deep learning algorithms to minimize ionizing radiation exposure during radiotherapy planning while preserving essential diagnostic information. We utilized CycleGAN+U-Net and SynDiff models to generate synthetic computed tomography (sCT) images from T1-weighted MRI scans of 37 patients exhibiting a spectrum of neurological conditions, ranging from mild cognitive impairment to severe Alzheimer’s disease. Model development and evaluation employed a rigorous k-fold cross-validation strategy. The dataset was partitioned, reserving 4 subjects as an independent test set. The remaining 33 subjects were utilized in a five-fold cross-validation process, systematically rotating subsets for training and validation to optimize model generalizability and performance assessment. From the comparison of the performances of the two applied models, SynDiff obtained higher segmentation accuracy values of ME (1066.92), MAE (1066.92), MSE (1474882.42), RMSE (1212.81), PSNR (34.67) and SSIM (0.001) than Cycle Gan + Unet and can be used to generate accurate sCT from brain MRI images. A key advantage of this AI-based sCT generation is the significant reduction in patient ionizing radiation exposure. This approach facilitates the preferential use of MRI, leveraging its superior soft-tissue contrast to provide essential information for detecting, monitoring, and diagnosing intracranial pathologies