Multi-contrast computed tomography healthy kidney atlas

Ho Hin Lee, Yucheng Tang, Kaiwen Xu, Shunxing Bao, Agnes B. Fogo, Raymond Harris, Mark P. de Caestecker, Mattias P. Heinrich, Jeffrey M. Spraggins, Yuankai Huo, Bennett A. Landman

TL;DR

This paper introduces the first multi-contrast CT atlas specifically for healthy kidneys, addressing the challenge of significant anatomical and contrast-based variability across individuals. By developing a novel deep learning-based cropping method and a two-stage registration pipeline, they create a standardized framework to integrate and visualize this variability, paving the way for better understanding of kidney anatomy across diverse populations and imaging protocols.

What Problem Does This Paper Address?

Understanding organ function and disease requires integrating information across multiple scales, from cellular to systemic. While significant effort has gone into mapping cellular and molecular details, contextualizing these findings within the broader organ and system level is crucial. The kidney, a retroperitoneal organ, presents a particular challenge for imaging due to substantial normal variations in its shape, size, and appearance, which are further complicated by different CT contrast phases (non-contrast, early arterial, late arterial, portal venous, delayed). Existing anatomical atlases are often limited in scope or modality, and there's a notable lack of a comprehensive atlas for abdominal and retroperitoneal organs that accounts for multi-contrast CT data. This gap hinders the ability to accurately analyze and interpret kidney anatomy across diverse patient populations and imaging conditions.

Methods

The authors propose a pipeline to construct a high-resolution CT retroperitoneal atlas optimized for kidneys across multiple contrast phases. The core of their approach involves:

  1. Deep Learning-Based Volume of Interest (VOI) Extraction: A deep neural network, termed "Body Part Regression" (BPR), is employed to automatically extract the abdominal and retroperitoneal regions from raw CT scans. This step is crucial for standardizing the field of view, reducing subjectivity, and improving the robustness of subsequent registration steps. The BPR network predicts a score for each slice, allowing for precise cropping to isolate the relevant anatomical areas.
  2. Two-Stage Hierarchical Registration Pipeline: After VOI extraction, a two-stage registration process is used to align the extracted volumes to a high-resolution atlas template.
  3. Affine Registration: This initial stage performs a global alignment, preserving 12 degrees of freedom to establish the overall spatial context.
  4. Non-Rigid Registration: This stage refines the alignment using a method called Dense Displacement Sampling (DEEDS), which leverages a self-similarity context metric. This metric focuses on local voxel-wise correspondence within patches, aiming to maintain anatomical integrity and handle deformations effectively.
  5. Atlas Generation and Evaluation: Successfully registered scans are averaged to create an "average map" and their variance is computed to generate a "variance map." These maps visualize the intensity fluctuations and anatomical variations across the population, providing insights into the stability and generalizability of the atlas.

Datasets

The study utilized two primary datasets:

  • Clinical Multi-Contrast Abdominal CT Cohort: This dataset comprised 2000 de-identified CT scans from patients aged 18-50, with criteria applied to select scans from individuals with healthy kidneys. After quality assessment, 720 subjects were included, providing a total of 290 unlabeled CT volumes across the five contrast phases (non-contrast, early arterial, late arterial, portal venous, delayed).
  • Multi-Organ Labeled Portal Venous Abdominal CT Cohort: A separate cohort of 100 portal venous contrast phase abdominal CT volumes was used. This dataset was valuable because it included ground truth labels for 13 different organs, allowing for comprehensive evaluation of the atlas's generalizability and the performance of different registration methods.

Key Insights and Ideas

  • The Importance of Multi-Contrast Atlases: The paper strongly emphasizes that a single-contrast atlas is insufficient for organs like the kidney, where contrast phases reveal distinct anatomical and functional details. A multi-contrast atlas is essential for a holistic understanding.
  • Deep Learning for Preprocessing: The integration of a deep learning model (BPR) for VOI extraction is a key innovation. It automates a critical, often manual, step, thereby increasing efficiency and reducing variability in the registration pipeline.
  • DEEDS Registration for Robustness: The choice of the DEEDS registration method, particularly its non-rigid component with a self-similarity metric, proved effective in handling anatomical variations and achieving stable registration for kidney organs.
  • Visualizing Variability is Key: The use of average and variance maps provides a powerful way to visualize not only the typical anatomy but also the extent of normal variation across a population, which is crucial for understanding disease states.

Key Results

  • DEEDS Outperforms Other Methods: In quantitative evaluations using the multi-organ labeled dataset, the DEEDS registration method, when integrated into their two-stage pipeline, consistently outperformed ANTS and NiftyReg across most organs, achieving a mean Dice score of 0.496 ± 0.284.
  • High Kidney Localization Accuracy: For the specific target organs (left and right kidneys), DEEDS demonstrated significantly better performance, with median Dice scores above 0.8. This indicates a high degree of accuracy in transferring kidney information to the atlas space.
  • Stable Deformation Across Sizes: Qualitative results, particularly the surface renderings with checkerboard patterns (Figure 5), showed stable deformation across a range of kidney volumetric sizes (100 cc to 308 cc), demonstrating the atlas's generalizability to morphological variations.
  • Variance Mapping Highlights Stable Regions: Variance maps (Figure 4) clearly showed low variability around the kidney regions, indicating stable registration and localization, while higher variability was observed in regions like the diaphragm, which are more prone to deformation.

Limitations and Future Work

  • Registration Method Optimization: While DEEDS performed well, the authors suggest that further optimization of the registration pipeline might be needed, especially considering the influence of neighboring organs like the liver and spleen.
  • Learning-Based Registration: They propose that learning-based registration methods, which have shown promise in brain imaging, could be explored for abdominal and retroperitoneal regions to potentially improve performance and reduce reliance on traditional similarity metrics.
  • Clinical Validation: The current atlas is based on healthy kidneys. Future work could involve clinical validation and phenotypic analysis to reveal biomarkers of specific organs in disease pathogenesis.
  • Broader Organ Coverage: The atlas framework could be extended to include segmentation of other abdominal and retroperitoneal organs, leveraging the multi-organ atlas label for improved segmentation models.

Summary updated

Abstract

The construction of three-dimensional multi-modal tissue maps provides an opportunity to spur interdisciplinary innovations across temporal and spatial scales through information integration. While the preponderance of effort is allocated to the cellular level and explore the changes in cell interactions and organizations, contextualizing findings within organs and systems is essential to visualize and interpret higher resolution linkage across scales. There is a substantial normal variation of kidney morphometry and appearance across body size, sex, and imaging protocols in abdominal computed tomography (CT). A volumetric atlas framework is needed to integrate and visualize the variability across scales. However, there is no abdominal and retroperitoneal organs atlas framework for multi-contrast CT. Hence, we proposed a high-resolution CT retroperitoneal atlas specifically optimized for the kidney across non-contrast CT and early arterial, late arterial, venous and delayed contrast enhanced CT. Briefly, we introduce a deep learning-based volume of interest extraction method and an automated two-stage hierarchal registration pipeline to register abdominal volumes to a high-resolution CT atlas template. To generate and evaluate the atlas, multi-contrast modality CT scans of 500 subjects (without reported history of renal disease, age: 15-50 years, 250 males & 250 females) were processed. We demonstrate a stable generalizability of the atlas template for integrating the normal kidney variation from small to large, across contrast modalities and populations with great variability of demographics. The linkage of atlas and demographics provided a better understanding of the variation of kidney anatomy across populations.
BibTeX
@article{lee2022multi,
  author = {Ho Hin Lee and Yucheng Tang and Kaiwen Xu and Shunxing Bao and Agnes B. Fogo and Raymond Harris and Mark P. de Caestecker and Mattias P. Heinrich and Jeffrey M. Spraggins and Yuankai Huo and Bennett A. Landman},
  title = {{Multi-contrast computed tomography healthy kidney atlas}},
  journal = {Computers in Biology and Medicine},
  year = {2022},
  doi = {10.1016/j.compbiomed.2022.105555},
}
Figure