Deep learning-based automated segmentation of eight brain anatomical regions using head CT images in PET/CT

Tong Wang, Haiqun Xing, Yige Li, Sicong Wang, Ling Liu, Fang Li, Hongli JingBMC Medical Imaging 2022 · DOI

TL;DR

This paper introduces a deep learning approach to automatically segment eight key brain regions from CT images acquired during PET/CT scans. By comparing their method to MRI-based segmentation, they demonstrate CT's potential for precise anatomical information, paving the way for more integrated PET/CT analyses.

What Problem Does This Paper Address?

Positron Emission Tomography/Computed Tomography (PET/CT) is a powerful tool for medical imaging, offering both functional (PET) and anatomical (CT) data. However, accurately segmenting specific brain regions from CT images, especially for subsequent functional analysis with PET data, has been challenging. While MRI is traditionally preferred for its superior soft tissue contrast and has been extensively studied for brain segmentation, it's not always feasible due to cost, patient contraindications (implants, claustrophobia), or emergency situations. This paper aims to bridge this gap by developing and evaluating a deep learning method for segmenting critical brain anatomical regions directly from CT scans, making PET/CT data more readily usable for detailed brain region analysis.

Methods

The researchers employed two popular deep learning architectures for semantic segmentation: DenseVNet and 3D U-Net. These Convolutional Neural Networks (CNNs) were trained on a dataset of 90 non-contrast CT (NCCT) scans. To establish a benchmark and validate their CT-based segmentation, they also performed segmentation on corresponding MRI scans using BrainSuite13, a well-established atlas-based method. The performance of the CNNs was primarily evaluated using the Dice similarity coefficient (DSC), a common metric for measuring segmentation overlap. Furthermore, they conducted a head-to-head comparison of the volumetric results obtained from their CT segmentation method and the MRI atlas method, assessing agreement using Intraclass Correlation (ICC) and correlation using Spearman's rank correlation coefficient.

Datasets

  • Training and Testing (CNNs): 90 patients with non-contrast computed tomography (NCCT) images were used. This dataset was split into 81 patients for training and 9 for testing the CNN models.
  • Independent Testing (CNNs) and Comparison: 18 patients who underwent both PET/CT and MRI were selected. Their CT scans were used as an independent testing set for the trained CNN models, and their MRI scans served as the reference for volumetric comparison.

Key Insights and Ideas

  • CT as a Viable Alternative for Brain Segmentation: The study demonstrates that with appropriate deep learning methods, CT images can provide precise anatomical information for brain segmentation, comparable to MRI in certain aspects. This is crucial for scenarios where MRI is not an option.
  • DenseVNet Outperforms 3D U-Net: For the specific task of segmenting these eight brain regions from CT, the DenseVNet architecture showed superior performance in terms of Dice scores compared to the 3D U-Net.
  • Region-Specific Performance: Segmentation accuracy varied significantly across different brain regions. While hemispheres and basal ganglia achieved high Dice scores, the hippocampus showed lower performance, suggesting that anatomical complexity and size play a role in segmentation difficulty.
  • Volumetric Agreement is Moderate: While Dice scores indicate good overlap for some regions, the agreement in voxel-based volumes between CT and MRI segmentation was only moderate for hemispheres and poor for other regions. This highlights the inherent differences between the modalities and potential challenges in direct volumetric comparisons.

Key Results

  • High Dice Scores for Key Regions: Four of the eight brain regions (left and right basal ganglia, left and right hemisphere) achieved high mean Dice scores (above 0.90) with the DenseVNet model on the independent testing set.
  • Moderate Agreement for Hemispheres: The Intraclass Correlation (ICC) for left and right hemispheres between CT and MRI volumes was moderate (0.618 and 0.654, respectively).
  • Moderate Correlation for Hemispheres: Spearman's correlation coefficients indicated moderate correlation for the left and right hemispheres (0.591 and 0.68, respectively) with MRI volumes.
  • Lower Performance for Hippocampus: The Dice scores for the left and right hippocampus were notably lower, particularly for the left hippocampus (0.089), indicating a significant challenge in segmenting these structures from CT.

Limitations and Future Work

  • Image Quality Differences: Variations in CT scanner parameters and post-processing between training and independent testing datasets might have influenced segmentation performance.
  • Dataset Size: The relatively small size of the training dataset could be a limiting factor for model generalization.
  • Region Size vs. Window Size: The spatial window size used for segmentation might have been too large for smaller structures like the hippocampus, potentially compromising their segmentation.
  • MRI Resolution: The MRI reference data had a relatively high slice thickness, which could have introduced inaccuracies in the atlas-based segmentation, particularly for fine anatomical regions.
  • Future Directions: The authors suggest further studies with larger datasets, varied spatial window sizes, and finer resolution MRI data. They also propose investigating the specific effects of imaging quality differences between CT and MRI. The work lays a foundation for future research with finer segmentation of regions and more complex anatomical investigations.

Summary updated

BibTeX
@article{wang2022deep,
  author = {Tong Wang and Haiqun Xing and Yige Li and Sicong Wang and Ling Liu and Fang Li and Hongli Jing},
  title = {{Deep learning-based automated segmentation of eight brain anatomical regions using head CT images in PET/CT}},
  journal = {BMC Medical Imaging},
  year = {2022},
  doi = {10.1186/s12880-022-00807-4},
}
Figure