Scholarlink

Brain tissue segmentation plays a fundamental role in clinical and research settings, creating detailed brain atlases, diagnosing neurological disorders such as dementia, and planning surgical interventions. Although magnetic resonance imaging is the preferred modality for its superior soft tissue contrast, CT serves as an accessible alternative to imaging in clinical routine and emergency situations. In this study, paired CT-MRI data sets from the Gothenburg H70 Birth Cohort (N=733) and the National University Hospital Memory Clinic Cohort, Singapore (NUS dementia cohort N=210) were used to train and evaluate advanced segmentation frameworks–nnUNet and MedNeXt–on CT brain segmentation guided by MRI-derived labels. In addition, the best performing models trained on Axial CT were selected to evaluate the sagittal and coronal CT. The 3D nnU-Net achieved average Dice Similarity Coefficients (DSCs) of 0.82, 0.72, and 0.76 for axial, coronal, and sagittal datasets, respectively, while MedNeXt exhibited slightly better performance with DSCs of 0.83, 0.73, and 0.78. MedNeXt demonstrated improved volumetric similarity across orientations, particularly in axial datasets, with scores ranging from 0.842 (CSF, sagittal) to 0.992 (WM, axial). For generalizability of the cohort to dementia patients, MedNeXt outperformed nnU-Net, achieving an average DSC and volumetric similarity of 0.73 and 0.912, compared to 0.70 and 0.854 for nnU-Net. Extended training (1000 epochs) improved the performance of nnUnet, while MedNeXt demonstrated superior scalability with increasing kernel sizes, requiring significantly longer training times (up to 288 h for large models). Thus, we demonstrate the higher training efficiency of nnUNet (requiring lower resource usage) and inference times, hence suitable for environments with limited resources, whereas MedNeXt consistently performs better and excels in multimodal imaging scenarios while maintaining clinically relevant accuracy across diverse datasets. Collectively, these findings validate automated brain tissue segmentation models for CT using paired MRI.

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

Brain Imaging and Tissue Segmentation from Paired CT-MRI Labels for Cognitively Normal and Dementia Cohort

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Vidya Somashekarappa,
  • Meera Srikrishna,
  • Silke Kern,
  • Joyce Chong Ruifen,
  • Eric Westman,
  • Christopher Chen,
  • Ingmar Skoog,
  • Jakob Seidlitz,
  • Michael Schöll

摘要

Brain tissue segmentation plays a fundamental role in clinical and research settings, creating detailed brain atlases, diagnosing neurological disorders such as dementia, and planning surgical interventions. Although magnetic resonance imaging is the preferred modality for its superior soft tissue contrast, CT serves as an accessible alternative to imaging in clinical routine and emergency situations. In this study, paired CT-MRI data sets from the Gothenburg H70 Birth Cohort (N=733) and the National University Hospital Memory Clinic Cohort, Singapore (NUS dementia cohort N=210) were used to train and evaluate advanced segmentation frameworks–nnUNet and MedNeXt–on CT brain segmentation guided by MRI-derived labels. In addition, the best performing models trained on Axial CT were selected to evaluate the sagittal and coronal CT. The 3D nnU-Net achieved average Dice Similarity Coefficients (DSCs) of 0.82, 0.72, and 0.76 for axial, coronal, and sagittal datasets, respectively, while MedNeXt exhibited slightly better performance with DSCs of 0.83, 0.73, and 0.78. MedNeXt demonstrated improved volumetric similarity across orientations, particularly in axial datasets, with scores ranging from 0.842 (CSF, sagittal) to 0.992 (WM, axial). For generalizability of the cohort to dementia patients, MedNeXt outperformed nnU-Net, achieving an average DSC and volumetric similarity of 0.73 and 0.912, compared to 0.70 and 0.854 for nnU-Net. Extended training (1000 epochs) improved the performance of nnUnet, while MedNeXt demonstrated superior scalability with increasing kernel sizes, requiring significantly longer training times (up to 288 h for large models). Thus, we demonstrate the higher training efficiency of nnUNet (requiring lower resource usage) and inference times, hence suitable for environments with limited resources, whereas MedNeXt consistently performs better and excels in multimodal imaging scenarios while maintaining clinically relevant accuracy across diverse datasets. Collectively, these findings validate automated brain tissue segmentation models for CT using paired MRI.