[go: up one dir, main page]
More Web Proxy on the site http://driver.im/ Skip to main content
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2021 Aug 4.
Published in final edited form as: Proc SPIE Int Soc Opt Eng. 2021 Feb 15;11596:115961T. doi: 10.1117/12.2580561

Construction of a Multi-Phase Contrast Computed Tomography Kidney Atlas

Ho Hin Lee a, Yucheng Tang a, Kaiwen Xu a, Shunxing Bao a, Agnes B Fogo b,c, Raymond Harris d, Mark P de Caestecker d, Mattias Heinrich e, Jeffrey M Spraggins f, Yuankai Huo a, Bennett A Landman a,g
PMCID: PMC8336653  NIHMSID: NIHMS1687683  PMID: 34354322

Abstract

The Human BioMolecular Atlas Program (HuBMAP) seeks to create a molecular atlas at the cellular level of the human body to spur interdisciplinary innovations across spatial and temporal scales. While the preponderance of effort is allocated towards cellular and molecular scale mapping, differentiating and contextualizing findings within tissues, organs and systems are essential for the HuBMAP efforts. The kidney is an initial organ target of HuBMAP, and constructing a framework (or atlas) for integrating information across scales is needed for visualizing and integrating information. However, there is no abdominal atlas currently available in the public domain. Substantial variation in healthy kidneys exists with sex, body size, and imaging protocols. With the integration of clinical archives for secondary research use, we are able to build atlases based on a diverse population and clinically relevant protocols. In this study, we created a computed tomography (CT) phase-specific atlas for the abdomen, which is optimized for the kidney organ. A two-stage registration pipeline was used by registering extracted abdominal volume of interest from body part regression, to a high-resolution CT. Affine and non-rigid registration were performed to all scans hierarchically. To generate and evaluate the atlas, multiphase CT scans of 500 control subjects (age: 15 – 50, 250 males, 250 females) are registered to the atlas target through the complete pipeline. The abdominal body and kidney registration are shown to be stable with the variance map computed from the result average template. Both left and right kidneys are substantially localized in the high-resolution target space, which successfully demonstrated the sharp details of its anatomical characteristics across each phase. We illustrated the applicability of the atlas template for integrating across normal kidney variation from 64 cm3 to 302 cm3.

Keywords: Computed Tomography, Kidney Atlas, Abdomen Atlas, Affine Registration, Non-Rigid Registration

1. INTRODUCTION

In the human body, trillions of cells are organized in tissues to perform parallel physiological processes [2]. The relationships between cells can be studied with multiple perspectives such as organization, specialization and cooperation. The Human Biomolecular Atlas Program (HuBMAP) from National Institute of Health (NIH) brings biologists, pathologists and bioinformaticians together with the main goal of studying and mapping the organization and molecular profiles of all cells within tissues or organs across the human body. While majority of efforts are distributed in the cellular and molecular perspectives, contextualizing and generalizing the anatomical characteristics of organs and systems in humans is crucial with the usage of computed tomography (CT). Creating a framework with the integration of micro-scale information and system-scale information benefits clinicians and researchers to visualize details across scales. The anatomical characteristics provided from the framework help linking the correspondence of the anatomical structures of organs to the pathogenesis of organ-related diseases, increasing the confidence level of the explainable findings across scales [3].

Within HuBMAP, the kidney is the organ that we are initially focusing on from the perspective of structural anatomy. Abdominal CT provides an opportunity to integrate information on the kidney at a system scale. Contrast enhancement CT is performed to emphasize the anatomical and structural information of organs and or neighboring vessels by injecting contrast agent before imaging procedures. Five contrast phases are generated corresponding to the remaining time of the contrast agent in the imaging cycle: 1) non-contrast, 2) early arterial, 3) late arterial, 4) portal venous and 5) delayed [4]. The relative intensity level within the kidney between each contrast phase CT varies to capture and specify the contextual characteristics of different organs. The structural and anatomical information of kidney can be clearly identified by integrating the details from multi-contrast phases. The atlas framework enables mapping kidney anatomical structure across variations in sex, body size (weight & height) and imaging protocols. Due to the large variability of anatomy and morphology of kidneys as shown in Figure 1., the creation of a standard reference template for abdominal organs is challenging and there is no abdominal atlas template currently available for public usage. An atlas reference template for kidneys can provide a framework to create a more in-depth understanding with its complex structure and anatomy localization across large clinical cohort.

Figure 1.

Figure 1.

Significant variations existed in the anatomical and morphological perspectives of kidneys across the demographics of large clinical cohorts. The creation of a standard atlas reference for abdominal organs remains challenging and hard to define a spatial domain which can share the contrast characteristics with large population of CT.

Atlases have widely been used in neuroimaging with the use of magnetic resonance imaging (MRI). Shi et al. proposed an infant brain atlas using unbiased group-wise registration with 3 different scanned time points (ie., neonate, 1 year old and 2 years old) of MRI from 56 males and 39 females normal infants [5]. Kuklisova-Murgasova et al. created atlases for early developing babies with age ranging from 29 to 44 weeks using affine registration [6]. Oishi et al. constructed a brain atlas with hierarchical affine and non-rigid registration with babies of 37–53 post-conceptional weeks [7]. Ali et al. developed an unbiased spatiotemporal 4-dimensional MRI atlas for the developing fetal brain by integrating symmetric diffeomorphic deformable registration with kernel regression in age, using normal fetuses scanned between 19 to 39 weeks of gestation and brain labeled structures [8]. Yuyao et al. proposed a novel framework to develop time-variable longitudinal atlas for infant brain in a spatial-temporal wavelet domain, using patch-based combination of results from each frequency subband [9]. Apart from generating early brain atlases for analyzing early development of brains, multiple modalities of adult and elderly atlases were also proposed to reveal the structural and contextual changes inside the brain for investigating neurological behavior and related diseases. James et al. derived two functional magnetic resonance imaging (fMRI) atlases from 21 healthy adults: 1) incorporating data from a single resting-state scan and 2) incorporating data from resting-state and task-based scans encompassing multiple perspectives such as visual perception, emotional processing and verbal memory[10]. Rajashekar et al. proposed two high-resolution normative brain atlases to investigate brain lesion-related diseases such as stroke and multiple sclerosis, across elderly population using FLAIR MRI and noncontrast CT [11]. These studies have shown great impacts on building brain atlases in clinical usage through multi-modality images. However, the development of atlases in abdomen is still remained challenging and few efforts have shown to be put into creating a standard reference in multi-modality (CT and MRI) images.

In this study, we constructed contrast-preserving CT abdominal atlases, optimizing for kidney organs with contrastcharacterized variability, and created an initial standard reference for multiple phases of CT in kidney organ perspective. In total of 500 healthy patients subjects with age ranging from 18-year-old to 50-year-old (250 males, 250 females) were used to compute the standard average template, to reveal the variability and biomarkers in kidneys across large population. Affine and non-rigid registration were performed hierarchically using DEEDS image registration tools [12, 13]. To our best knowledge, this is the first attempt of creating multi-contrast phases CT abdominal kidney atlases for publicity usage.

2. METHODS

In this study, a two-stage registration pipeline was created for transferring the spatial information of abdominal organs to the standard reference atlas target space, across large population of clinical cohort, which is presented in Figure 2. As described below, this process was optimized for localization of kidneys with multi-contrast characteristics.

Figure 2.

Figure 2.

The complete pipeline for creating the standard reference template of kidney can be divided into 3 steps: 1) Preprocessing, 2) Performing registration and 3) Calculating average. The abdominal volume of interest is extracted with the algorithm body part regression, using slice information to classify the specific location of where the slice is in the body [1]. The volume interest is then resampled to the same resolution of the high-resolution atlas target. The resampled output is registered to the atlas target with affine and non-rigid registration hierarchically. All successfully registered volumes are finally used to generate an unweighted average template, localizing the anatomical information of kidney in each contrast phase CT.

2.1. Preprocessing

3D abdominal multi-contrast CT volumes are used as the input for the registration pipeline. However, the difference in the field of view within the large population of clinical scans is significant, leading to a certain amount of failure registrations. An automatic deep learning based algorithm body part regression is introduced and help extract the abdominal volume of interest by using slice information [1]. 2D slice images are extracted from the whole 3D volume as the input for the body part regression algorithm and a predicted value is computed for each slice as the output. Each output value from the algorithms represented the particular slice’s approximate location in the whole volume of 3D images. The output value for each slice is in the range of −12 to +12, which starts from the location near the heart to the location around pelvis respectively. Slices with predicted value −5 to +5 are empirically extracted and stacked as 3D volume for the abdominal volume of interest. The extracted volumes are then resampled to the same resolution of the atlas target with padding or cropping to maintain the same shape with the atlas target for the registration pipeline.

2.2. Registration Pipeline

The two-stage registration pipeline is introduced using dense displacement sampling (DEEDS), which is a 3D medical registration tool based on a discretized search space [12, 13]. The image similarity for a number of randomly sampling voxels of each control point is calculated and models the incremental diffusion regularization through a pair-wise term for the deformation field. The complete pipeline is illustrated in Figure 2, which can be divided into: 1) DEEDS affine registration and 2) DEEDS non-rigid registration. The extracted volume of interest from all scans had the same resolution and dimensions with the atlas target after the preprocessing step. Affine registration was initially performed to handle multiple independent movements by assigning prior definition of each affine component and its spatial context. The nonrigid registration was then initialized using the output of affine registration based on the same similarity metric, which illustrated in the later section, with a similar block-matching search and trimmed least squares. Five scale levels are computed in DEEDS with grid spacing ranging from 8 to 4 voxels, displacement search radii from six to two steps between 5 and 1 voxels [1214]. A pair of deformed scan and displacement data file for all sampled control points is generated as the output of DEEDS. The deformed scan is in the same spatial space with the atlas target after registration pipeline. The average template of each contrast phase is computed using all successful deformed results.

2.3. Similarity Metric for Registration

The similarity metric of DEEDS registration tool is to measure the patch-based self-similarity context and aims to search the context around the voxel of the extracted volume interest [15]. Measuring the self-similarity can be demonstrated as optimizing a distance function M between image patches extracted from the original image I, an estimation function n2 of both local and global noise, and a certain neighborhood layout N for determining the kinds of self-similarities. By randomly extracting a patch centered at a, the self-similarity measurement can be described as following:

M(I,a,b)=exp(S(a,b)n2)a,bN (1)

where b defined the center location of a patch from one of the neighborhood N. The advantage of using this similarity metric is to avoid the disadvantage of image artifacts or noise from the central patch and prevent a direct adverse effect in calculation. The pairwise distance between patches is calculated within 6 neighborhoods and compare the contextual information within the neighborhood, instead of abstracting good shape representation.

3. DATA AND EXPERIMENTS

3.1. Data and Platform

Initially, 2000 patients’ abdomen CT data were retrieved de-identified form from ImageVU under the Institutional Review Board (IRB) approval. Then, we used exclusion criteria for identifying normal kidneys from subjects. Out of 2000, 720 subjects are retrieved after assessment by standard ICD-9 codes and age 18–50 years old. Out of 720, we limited our study to those who had specific contrast phase abdominal CT scans, which included total of 290 CT volumes (non-contrast: 50 volumes, early arterial: 30 volumes, late arterial: 80 volumes, portal venous: 100 volumes, delayed: 30 volumes). The atlas target is a single subject volume with dimension of 512 × 512 × 434. It is chosen with high resolution and significant contrastive and morphological characteristics in kidney organs. All volumes were resampled to the same resolution and dimensions of the atlas target, in which the resolution and the dimensions of the atlas target. The resampled outputs were then used as the input for the registration pipeline to transfer the spatial information towards the atlas target space.

3.2. Atlas Construction

3.2.1. Affine Registration

All 3D volumes were first affine aligned with the atlas target with 12 degrees of freedom. A transformation 4-by-4 matrix was identified for each affine aligned scan in the target space. The output transformation matrix for each affine aligned scan was computed according to the five hierarchical levels with displacement search radii from six to two steps between 5 and 1 voxels iteratively, selecting the optimal displacement for each control point as pre-alignment using least trimmed squares.

3.2.2. Non-Rigid Registration

To initialize the non-rigid registration, the affine registered output was the input in the format of an affine transformation matrix. The original 3D volume was then transferred to the affine registered space. The non-rigid registration was performed with the affine registered input by approximating the neighborhood relations of the control point grid using a minimum-spanning-tree, extracting a smooth displacement field [14, 15]. However, the non-rigid registration process is highly sensitive to the field of view between the moving image and the atlas target image. The significant variability of the field of view between patches leads to a great adverse effect of calculating the neighborhood relations and generate excess deformation on the bottom/top of the body. Therefore, we performed visualization manually to the registered results and chose successfully registered output for calculation of the atlas average template. Two conditions were applied for choosing the good registered result: 1) Good registration and align to the abdominal body of the atlas target, and 2) Kidney are stably localized in a similar anatomical location with that of atlas target. The average atlas template and the variance map of multi-phases were then generated with the well-registered outputs.

4. RESULTS

The generated results from single subject registration were first evaluated by overlaying the registered output to the atlas target and manually visualizing the quality of the body alignment between both images. The single subject registered output of each contrast phase is demonstrated in the upper block of Figure 3. The abdominal body of each phase’s registered output is stable and aligned well to the body of atlas target by overlaying to each other. The main focus for this abdominal atlas is to optimize for kidney organ and evaluate the ability to localize the kidney organ from 3D volumes of each phase. Both left and right kidney were substantially localized in the similar anatomical location compared to that of the atlas target and deformed to the similar anatomical structure in the atlas target space. All successfully registered volumes in the atlas space were then summed to compute the average atlas templates, which were shown in the lower block of Figure 3. The contrast characteristics of each phase were transferred successfully to the atlas space and the variability of the intensity in kidney was demonstrated across large population of clinical cohorts.

Figure 3.

Figure 3.

The qualitative comparison of the single subject registration and the computed average templates across multi-contrast phases are shown here. Significant differences of the voxel intensity in both left and right kidneys are demonstrated corresponding to the specific contrast imaging protocols. Kidneys from all multi-phases average templates are localized with detailed structural information in the atlas space and the contrast characteristics of each phase are well allocated in kidneys.

While the average template of each phase provided significant approval of how the intensity-based information is successfully transferred and optimized in kidneys, the structural and contextual information are also important perspectives to show the stability of the average template computed from such large population cohort. Variance maps were computed from the average template to visualize the variability of multiple organs throughout all subjects in the clinical cohort. The variance value of each voxel was calculated from the intensity variation contributed to the average template with each subject data. In Figure 4, low variance value was shown in both left and right kidney, while a high variance value was observed around the liver and spleen region.

Figure 4.

Figure 4.

Variance map are generated and small variation of intensity is shown in both left and right kidney. The kidneys from each subjects’ scans are well registered to the atlas target with small variability. The value of the variance map represents the log of the variation in voxel intensity.

Meanwhile, an inverse transformation from atlas space to original image space was also performed to see the functionality of the atlas transferring the information back to the original space. 3D volume of both left and right kidneys were first generated for each subject data and a checkerboard pattern grid from atlas space was created in the coronal view of the volume. The inverse transformation was then applied to the checkerboard pattern gird and to see the stability of the deformation to transfer back to original spatial space for kidneys. From Figure 5, the checkerboard pattern grid in 3D volume rendering demonstrated a stable transfer from the atlas space and deformed towards similar structural characteristics of kidneys in the original space label, while the deformation is significant near the boundary of the kidneys.

Figure 5.

Figure 5.

3D volume rendering was performed to both left and right kidneys with deformed checkerboard pattern from atlas space (Red: right kidney original space label, blue: left kidney original space label, yellow: atlas inversely transformed to original space checkerboard pattern), where 1 cc is equal to 1 cubic centimeter. The deformed pattern showed stable deformation result using inverse transform to obtain the original image space information from atlas target.

5. DISCUSSION AND CONCLUSION

From the result shown above, the contrast characteristics of each contrast phase were successfully demonstrated in the atlas average template and low variance intensity is shown around the anatomical location of kidney in the atlas space. However, high variance of intensity was observed in both spleen and liver region and it located the large structural variability of both livers and spleen organs across all subjects in the clinical cohort. Body part regression provided an opportunity to extract approximate regional interest for abdomen, while the difference in field of view from original volumes may contribute to high variability in the upper region of abdomen. DEEDs demonstrated a stable performance on registering volumes to the atlas space, while it is highly sensitive to the variation of field of view. The significant difference of field of view may contribute to a large extent of deformation and lead to failure registrations. In conclusion, we constructed a stable standard reference template for abdominal organs and well localized the contextual information of kidneys with multi-phases characteristics of CT. The low variation of intensity in kidneys shows a stable registration of kidneys across the clinical cohorts from moving image space to atlas space. The high variance of intensity around other organs is a potential area for improvement in future perspectives.

6. ACKNOWLEDGEMENTS

This research is supported by NIH Common Fund and National Institute of Diabetes, Digestive and Kidney Diseases U54DK120058 (Spraggins), NSF CAREER 1452485, NIH 2R01EB006136, NIH 1R01EB017230 (Landman), and NIH R01NS09529. This study was in part using the resources of the Advanced Computing Center for Research and Education (ACCRE) at Vanderbilt University, Nashville, TN. The identified datasets used for the analysis described were obtained from the Research Derivative (RD), database of clinical and related data. The imaging dataset(s) used for the analysis described were obtained from ImageVU, a research repository of medical imaging data and image-related metadata. ImageVU and RD are supported by the VICTR CTSA award (ULTR000445 from NCATS/NIH) and Vanderbilt University Medical Center institutional funding. ImageVU pilot work was also funded by PCORI (contract CDRN-1306-04869).

REFERENCES

  • [1].Yan K, Lu L, and Summers RM, “Unsupervised body part regression via spatially self-ordering convolutional neural networks,” in 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018), 2018: IEEE, pp. 1022–1025. [Google Scholar]
  • [2].Rozenblatt-Rosen O, Stubbington MJ, Regev A, and Teichmann SA, “The Human Cell Atlas: from vision to reality,” Nature News, vol. 550, no. 7677, p. 451, 2017. [DOI] [PubMed] [Google Scholar]
  • [3].Consortium H, “The human body at cellular resolution: the NIH Human Biomolecular Atlas Program,” Nature, vol. 574, no. 7777, p. 187, 2019. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [4].Donath T et al. , “Toward clinical X-ray phase-contrast CT: demonstration of enhanced soft-tissue contrast in human specimen,” Investigative radiology, vol. 45, no. 7, pp. 445–452, 2010. [DOI] [PubMed] [Google Scholar]
  • [5].Shi F et al. , “Infant brain atlases from neonates to 1-and 2-year-olds,” PloS one, vol. 6, no. 4, p. e18746, 2011. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [6].Kuklisova-Murgasova M et al. , “A dynamic 4D probabilistic atlas of the developing brain,” NeuroImage, vol. 54, no. 4, pp. 2750–2763, 2011. [DOI] [PubMed] [Google Scholar]
  • [7].Oishi K et al. , “Multi-contrast human neonatal brain atlas: application to normal neonate development analysis,” Neuroimage, vol. 56, no. 1, pp. 8–20, 2011. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [8].Gholipour A et al. , “A normative spatiotemporal MRI atlas of the fetal brain for automatic segmentation and analysis of early brain growth,” Scientific reports, vol. 7, no. 1, pp. 1–13, 2017. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [9].Zhang Y, Shi F, Wu G, Wang L, Yap P-T, and Shen D, “Consistent spatial-temporal longitudinal atlas construction for developing infant brains,” IEEE transactions on medical imaging, vol. 35, no. 12, pp. 2568–2577, 2016. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [10].James GA, Hazaroglu O, and Bush KA, “A human brain atlas derived via n-cut parcellation of resting-state and task-based fMRI data,” Magnetic resonance imaging, vol. 34, no. 2, pp. 209–218, 2016. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [11].Rajashekar D et al. , “High-resolution T2-FLAIR and non-contrast CT brain atlas of the elderly,” Scientific Data, vol. 7, no. 1, pp. 1-7, 2020. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [12].Heinrich MP, Jenkinson M, Brady M, and Schnabel JA, “MRF-based deformable registration and ventilation estimation of lung CT,” IEEE transactions on medical imaging, vol. 32, no. 7, pp. 1239–1248, 2013. [DOI] [PubMed] [Google Scholar]
  • [13].Heinrich MP, Maier O, and Handels H, “Multi-modal Multi-Atlas Segmentation using Discrete Optimisation and Self-Similarities,” VISCERAL Challenge@ ISBI, vol. 1390, p. 27, 2015. [Google Scholar]
  • [14].Xu Z et al. , “Evaluation of six registration methods for the human abdomen on clinically acquired CT,” IEEE Transactions on Biomedical Engineering, vol. 63, no. 8, pp. 1563–1572, 2016. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [15].Heinrich MP, Jenkinson M, Papież BW, Brady M, and Schnabel JA, “Towards realtime multimodal fusion for image-guided interventions using self-similarities,” in International conference on medical image computing and computer-assisted intervention, 2013: Springer, pp. 187–194. [DOI] [PubMed] [Google Scholar]

RESOURCES