key: cord-0503087-os8vodmz
authors: Yang, Zhenyu; Lafata, Kyle J; Chen, Xinru; Bowsher, James; Chang, Yushi; Wang, Chunhao; Yin, Fang-Fang
title: Quantification of lung function on CT images based on pulmonary radiomic filtering
date: 2021-05-24
journal: nan
DOI: nan
sha: 745ff56ac6640a9575f69436bf70a7af9946c24c
doc_id: 503087
cord_uid: os8vodmz

Purpose: To develop a radiomics filtering technique for characterizing spatial-encoded regional pulmonary ventilation information on lung CT. Methods: The lung volume was segmented on 46 CT images, and a 3D sliding window kernel was implemented across the lung volume to capture the spatial-encoded image information. Fifty-three radiomic features were extracted within the kernel, resulting in a 4th order tensor object. As such, each voxel coordinate of the original lung was represented as a 53-dimensional feature vector, such that radiomic features could be viewed as feature maps within the lungs. To test the technique as a potential pulmonary ventilation biomarker, the radiomic feature maps were compared to paired functional images (Galligas PET or DTPA-SPECT) based on Spearman correlation (r) analysis. Results: The radiomic feature map GLRLM-based Run-Length Non-Uniformity and GLCOM-based Sum Average are found to be highly correlated with the functional imaging. The achieved r (median [range]) for the two features are 0.46 [0.05, 0.67] and 0.45 [0.21, 0.65] across 46 patients and 2 functional imaging modalities, respectively. Conclusions: The results provide evidence that local regions of sparsely encoded homogenous lung parenchyma on CT are associated with diminished radiotracer uptake and measured lung ventilation defects on PET/SPECT imaging. These findings demonstrate the potential of radiomics to serve as a complementary tool to the current lung quantification techniques and provide hypothesis-generating data for future studies.

Lung disease is a leading cause of global mortality [1, 2] , and assessment of pulmonary function is essential in patients with lung disease [3, 4] . Ventilation denotes the exchange of blood and gas in the lungs and is a common surrogate of lung function [3] . The identification of the ventilation defects helps screen lung diseases [4] , evaluate radiation therapy treatment plans [5] , and minimize treatment-induced parenchymal injury [6] . Over the past few decades, nuclear medicine-based ventilation scintigraphy has served as the clinical gold standard for tomographic pulmonary function imaging and assessment [7] . Typically, nuclear medicine-based scintigraphy is costly, and prophylactic applications are limited due to commonly observed adverse side effects [8] . Such limitations, along with the growing socioeconomic burdens linked to lung disease, have motivated the development of new quantitative imaging techniques in recent years [9] .

Advances in new techniques to characterize high-dimensional feature spaces [10] and quantitative applications of CT imagingwhich is currently the standard-of-care for lung imaging -suggest that lung disease and pulmonary function can be treated as a regionally heterogeneous system [9, 11, 12] . Earlier efforts in quantitative pulmonary function assessment include Hounsfield Unit (HU)-based thresholding, but the reported performance was limited [12] . Computed tomography ventilation imaging (CTVI) is an emerging pulmonary function quantification method using respiratory-gated four-dimensional CT (4DCT).

In some reported studies, deformable image registration (DIR) and motion modeling analysis [13] [14] [15] were used in CTVI to characterize breathing-induced changes of lung density and air volume [3, 9, 14] . Such procedure not only relies on the acquisition of 4DCT, but also its results can be sensitive to the choice of DIR algorithm and motion model [5, [16] [17] [18] .

Radiomics is a popular medical imaging quantitative framework, which employs feature engineering techniques to convert standard-of-care medical images into mineable data [19] . Previous radiomic-based pulmonary studies have characterized the intensity, texture, and morphology of the lungs and shown association with global pulmonary function measurements, e.g., forced expiratory volume in one second (FEV1) [11] . However, radiomic analysis of the total lung volume as a region-of-interest (ROI) is insufficient to characterize spatial-encoded pathological and physiological characteristics of the lungs.

Spatial information is critical for the clinical interpretation of lung function, e.g., in the preoperative assessment of lung resection candidates with impaired lung function [20] . The sub-regional radiomic analysis, known as habitat analysis, combined with machine learning and deep learning, has been reported to identify lung function [21] [22] [23] . However, the direct analysis of true spatial-encoded radiomic features with the regional pulmonary ventilation is needed.

In this work, a radiomics filtering technique is developed to capture spatial-encoded image pathological and physiological features associated with underlying ventilation changes from lung CT. Motivated by habitat radiomics, the developed radiomic filtering technique extends the radiomic extraction to a voxel basis. To test the radiomic filtering of the lungs, we applied our technique to a publicly available CT dataset, where nuclear medicine-based ventilation images were also available as reference measurements. The developed technique may serve as a complementary tool to current pulmonary quantification techniques and provides hypothesis-generating data for future studies in quantitative pulmonary imaging.

The following methods were carried out under relevant regulations. Retrospective data analysis was completed with approval from the Duke University Health System Institutional Review Board. All data is de-identified and is publicly available as part of the VAMPIRE Challenge dataset [3] .

A public non-small cell lung cancer (NSCLC) patient dataset from the VAMPIRE Challenge [3] was used as the basis of this work. Specifically, the VAMPIRE dataset includes paired image acquisitions of CT and reference ventilation images (RefVI) crossing ventilation imaging modalities and clinical settings [3] . Average CT volumes were derived for all 46 patients and finally adopted for our radiomic analysis. All

RefVIs (PET and SPECT) were obtained after CT acquisition and were co-registered to the corresponding average CT image. For each patient, lung segmentation masks were included in the dataset for both the average CT image and the RefVI image. 

The overall research design and methodological workflow, including radiomic filtering, feature map analysis, and functional imaging comparison, is summarized in Figure 1 . The lung volume was first obtained from the average CT image and the segmentation mask for each patient. A 3D sliding window kernel was implemented to capture the regional radiomic features throughout the entire lung volume.

Accordingly, each voxel coordinate in the original lung was represented as a -dimensional feature vector, and the feature space can be represented as a -dimensional 4 th order tensor, i.e., a set of shift-invariant 3D feature maps. Due to this shift-invariance property, each feature map can be viewed in the same reference frame as the original CT image. Formally, given radiomic features and a CT image with a dimension of × × , the feature maps can be mathematically represented as a multi-dimensional tensor:

where the ( , , , ) th component denotes the measured value of the th radiomic feature at the ( , , ) th tomographic coordinate.

In this work, a total of = 53 radiomic features were included. These features were chosen to collectively capture the local texture intensity and texture characteristics within the lungs [19] . Table I 

The radiomic filtering was performed with respect to the 4DCT average lung volume. Following the previous lung radiomic studies, the dynamic range of the lung was re-binned to 64 gray levels [11, 24, 25] .

The choice of sliding window (kernel) size defines the radiomic filtering resolution. A small kernel may lead to a noisy output feature map, while a large kernel may lose the signal. The typical pulmonary defect volumes vary from 20 cm 3 to 80 cm 3 , and the diameter is 20-40 mm [26, 27] . Here, we set a kernel size of 15 × 15 × 15 mm 3 , which is on the order of a typical pulmonary defect. All 53 features were averaged over 13 directions [24] to approximate rotational invariance [28] [29] [30] [31] .

All the calculations were performed using an in-house developed radiomics filtering toolbox with MATLAB (MATLAB R2018a; MathWorks, Natick, Mass). The toolbox was fully validated against the image biomarker standardization initiative (IBSI) standardization [32] as well as the digital phantoms [33] .

In addition, the toolbox was specifically optimized for this voxel-based, rotationally invariant, 3D

calculation. It included (1) fully vectorized calculation procedures to accelerate small-kernel matrix manipulation and (2) parallel computing to accelerate distributed computation. The average calculation time for each patient was about 10 minutes, which is effective considering that the operation was performed on the entire 3D lung volume at native resolution.

To reduce feature redundancy in equation (1), we performed a dimensionality reduction procedure as follows. For each patient, Pearson correlation coefficients (r) of each feature pair were calculated, thereby leading to 46 covariance matrices. The mean Pearson ̅ covariance matrix was considered as a representation of the multi-collinearity existing in , and the feature maps with ̅ > 0.95 were considered as highly correlated [19, 34, 35] . The highly correlated features were classified as a feature subset and treated as a group in the subsequent analysis.

The extracted radiomic feature maps were quantitatively compared to the corresponding RefVI. For each feature map/RefVI matched pair, voxel-wise Spearman correlation coefficients ( ) were calculated within the lungs. Prior to the calculation, all feature maps were resampled to match the spatial dimensions of the RefVI via a nearest-neighbor interpolation. To keep consistency with the VAMPIRE study, the RefVI images were smoothed as follows: voxels with ventilation intensity above ±4 standard deviation of the mean intensity of overall RefVI lung voxels were removed until the threshold converged to within 1% of the last threshold [3, 36] .

We hypothesized that the joint measurement of image intensity and texture by radiomics analysis would provide complementary information compared to conventional intensity-only (e.g., thresholding) analysis [25] . To test the hypothesis, we evaluated the following three intensity-based studies:

1) The Spearman correlation between original lung CT with RefVI.

2) The Spearman correlation between average filtering lung CT with RefVI. The average filter calculates the regional mean within a 15 × 15 × 15 mm 3 sliding window kernel for objective comparison with radiomics filtering.

3) The regional overlap between the pulmonary defects was identified by intensity thresholding with ground truth defects. The CT value of −950 HU is considered as an acceptable threshold between emphysema and normal lung [12] . According to the VAMPIRE study, the region with the lowest 25% percent of the total intensity in PET/SPECT is considered as the ground truth pulmonary defects [3] . The Dice coefficient was employed to quantify the regional overlap.

The Spearman correlations from the first and second studies were evaluated against the highest-performing radiomic results. Wilcoxon signed-rank tests were performed on the Spearman coefficient, where p<0.05 was considered statistically significant. 

The assessment of regional pulmonary function is essential in patients with lung disease. Modern imaging systemsin combination with state-of-the-art computational techniques (e.g., radiomics)enable new approaches to this problem. In this work, we hypothesize that spatially encoded radiomic textures are inherently associated with local changes in pulmonary function. To test the hypothesis, we developed a radiomic filtering workflow to characterize the association between regionally heterogeneous lung CT image [37] . The HU value in a certain lung region can be considered as a linear combination of a waterlike material and air-like material [38] [39] [40] . Hence, the GLCOM-based Sum Average can be taken as a measurement of the local air content change [41, 42] . It is generally known that the reduction of pulmonary function is closely associated with diffuse alveolar epithelium destruction, capillary damage/bleeding, hyaline membrane formation, alveolar septal fibrous proliferation, and pulmonary consolidation [43] [44] [45] .

The reduction in lung function can be assumed as the healthy pulmonary tissue is replaced by an increase of air within the lungs [25] . As such, the reduction in lung function leads to regional heterogeneity and a decrease in CT attenuation coefficient, which could be potentially reflected as lower GLRLM-based Run-Length Non-Uniformity and lower GLCOM-based Sum Average. Collectively, these two features suggest that local regions of sparsely encoded homogenous lung parenchyma on CT are associated with diminished radiotracer uptake and measured lung ventilation defects on PET/SPECT imaging. This finding is consistent with previous whole lung-based analysis of emphysema [46] , asthma [47] , lung inflammation [48] , COPD [49] , fibrosis [50].

The VAMPIRE Challenge benchmarked 37 independent CTVI algorithms using the VAMPIRE dataset based on the Spearman coefficient [3] . Since this work employed the same dataset as well as the statistical Although this research has demonstrated promising results and has provided hypothesis-generating data, our study also has several limitations. First, the radiomic filtering technique may also capture the nonparenchymal voxel intensities (e.g., blood vessels and artifacts). The VAMPIRE dataset has excluded the major blood vessels from the lung contours. The pulmonary vessels with a cross-sectional diameter < 5 mm are also shown to be related to emphysema and pulmonary hypertension [52, 53] . We hypothesized that the small blood vessels could also be regarded as a regional heterogeneous system [54] , closely related to lung ventilation and airflow, and captured by radiomics. We note that such vessel segmentation masks, as provided by the VAMPIRE, also largely preclude the drawbacks of DIR-based CTVI (e.g., extended processing time). The incorporation of DIR into image fusion software and radiotherapy treatment planning systems also reduces the need for expert users. Quantification of radiomic filtering in the pulmonary vasculature image can be considered as a future research direction. Second, the current radiomic filtering model was designed and implemented in 4DCT average volume. 4DCT average volume has been reported to be highly consistent with free-breathing CT in radiotherapy contouring, planning, and dose calculation for lung cancer patients [55] . However, the sorting artifacts caused by lung motion may lead to locally unresolvable uncertainties in the radiomics filtering feature maps. Further characterization on the impact of 4DCT averaged images, phase images, free-breathing images, and breath-hold images on voxel-wise feature quantification are warranted, particularly to further investigate minimum resolvable texture. Due to the limitation of the current VAMPIRE dataset, the performance of radiomic filtering on free-breathing CT remains unknown. It is worth exploring the free-breathing CT scans, which could potentially expand the utility and reach of the technique. Finally, the variability of possible CT acquisition settings can also introduce uncertainties for the radiomic analysis. Typically, the radiomic analysis is sensitive to institution-15 specific differences in kV tube settings (kVp/mAs), scan collimation/slice thickness, consistency of HU calibration, choice of standard or iterative reconstruction, and the prevalence of 4DCT motion artifacts (e.g., different types of phase/amplitude binning, and selection of scan pitch) [56] [57] [58] [59] [60] . In this study, the fully curated VAMPIRE dataset has relatively uniform acquisition parameters, signal-to-noise ratio, etc.

The filtering process focuses on the relative value of the feature in a regional kernel, rather than its absolute values, which may be less sensitive than the traditional radiomic approach. However, a more comprehensive generalization and robustness analysis of our technique should be further characterized as part of future work.

In this work, we presented a computational radiomic filtering technique to quantify local changes in pulmonary function on CT images. Our results provide evidence that local regions of sparsely encoded homogenous lung parenchyma on CT are associated with diminished radiotracer uptake and measured lung ventilation defects on PET/SPECT imaging. This finding demonstrates the potential of radiomic filtering to provide hypothesis-generating data for future studies, and could eventually serve as a clinically useful tool to potentially help diagnose and screen for lung disease.

All imaging data used in this study is obtained from the publicly available VAMPIRE dataset [3] . The original data sets can be downloaded via request from the VAMPIRE website: https://imagex.sydney.edu.au/vampire-challenge/.

Deaths: leading causes for 2015

The global burden of respiratory disease-impact on child health

The VAMPIRE challenge: A multi-institutional validation study of CT ventilation imaging. Medical physics

Gas exchange and ventilation-perfusion relationships in the lung

Imaging of regional ventilation: Is CT ventilation imaging the answer? A systematic review of the validation data

Functional lung imaging: emerging methods to visualize regional pulmonary physiology1

68Ga PET/CT ventilation-perfusion imaging for pulmonary embolism: a pilot study with comparison to conventional scintigraphy

Quantification of heterogeneity in lung disease with image-based pulmonary function testing

What can computed tomography and magnetic resonance imaging tell us about ventilation

Data clustering based on Langevin annealing with a self-consistent potential

An exploratory radiomics approach to quantifying pulmonary function in CT images

Normal Lung Quantification in Usual Interstitial Pneumonia Pattern: The Impact of Threshold-based Volumetric CT Analysis for the Staging of Idiopathic Pulmonary Fibrosis

New pathways for end-to-end validation of CT ventilation imaging (CTVI) using deformable image registration

Pulmonary ventilation imaging based on 4-dimensional computed tomography: comparison with pulmonary function tests and SPECT ventilation images

Clinical validation of 4-dimensional computed tomography ventilation with pulmonary function test data. International journal of radiation oncology, biology, physics

CT ventilation imaging derived from breath hold CT exhibits good regional accuracy with Galligas PET

Quantifying the reproducibility of lung ventilation images between 4-Dimensional Cone Beam CT and 4-Dimensional CT

Measuring interfraction and intrafraction lung function changes during radiation therapy using four-dimensional cone beam CT ventilation imaging

Radiomics: images are more than pictures, they are data

Assessing regional lung function. Thoracic Surgery Clinics

Chronic obstructive pulmonary disease: thoracic CT texture analysis and machine learning to predict pulmonary ventilation

Pulmonary ventilation maps generated with free-breathing proton MRI and a deep convolutional neural network

A deep learning method for producing ventilation images from 4DCT: First comparison with technegas SPECT ventilation

Spatial-temporal variability of radiomic features and its effect on the classification of lung cancer histology

Novel Identification of Radiomic Biomarkers with Langevin Annealing

Hyperpolarized 3He Magnetic Resonance Imaging of Ventilation Defects in Healthy Elderly Volunteers

Evaluation of Asthma With Hyperpolarized Helium-3 MRI. Chest

Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach

Texture information in run-length matrices

Prediction of clinical phenotypes in invasive breast carcinomas from the integration of radiomics and genomics data

Radiomics: a primer on high-throughput image phenotyping

The Image Biomarker Standardization Initiative: Standardized Quantitative Radiomics for High-Throughput Image-based Phenotyping

Digital phantoms for characterizing inconsistencies among radiomics extraction toolboxes

Test-retest reproducibility analysis of lung CT image features

Differences in patient outcomes of prevalence, interval, and screendetected lung cancers in the CT arm of the National Lung Screening Trial

202-04: The VAMPIRE Challenge: Preliminary Results From a Multi-Institutional Study of CT Ventilation Image Accuracy. Medical physics

Investigation of the cutting conditions in milling operations using image texture features

Dynamic ventilation imaging from four-dimensional computed tomography

Non-invasive imaging of regional lung function using x-ray computed tomography

Quantification of regional ventilation from treatment planning CT

A novel and fully automated mammographic texture analysis for risk prediction: results from two case-control studies

CT-based ventilation imaging in radiation oncology

Abnormal pulmonary function in COVID-19 patients at time of hospital discharge

The long-term impact of severe acute respiratory syndrome on pulmonary function, exercise capacity and health status

The 1-year impact of severe acute respiratory syndrome on pulmonary function, exercise capacity, and quality of life in a cohort of survivors

Utility of lung density measurements in the diagnosis of emphysema. Respiratory medicine

Decreased computed tomographic lung density during exacerbation of asthma

Lung density on high resolution computer tomography (HRCT) reflects degree of inflammation in smokers

The human microbiome in the lung: are infections contributing to lung health and disease? Chronic Obstructive Pulmonary Diseases

Estimating lung ventilation directly from 4D CT Hounsfield unit values. Medical physics

CT pulmonary vessels and MRI ventilation in chronic obstructive pulmonary disease: relationship with worsening FEV1 in the TINCan cohort study

Quantitative CT measurement of cross-sectional area of small pulmonary vessel in COPD: correlations with emphysema and airflow limitation

Distribution of blood flow and ventilation in the lung: gravity is not the only factor. British journal of anaesthesia

Comparison of CT images with average intensity projection, free breathing, and mid-ventilation for dose calculation in lung cancer

Comprehensive investigation on controlling for CT imaging variabilities in radiomics studies

Intrinsic dependencies of CT radiomic features on voxel size and number of gray levels. Medical physics

Assessing agreement between radiomic features computed for multiple CT imaging settings

Bladder cancer treatment response assessment in CT using radiomics with deeplearning. Scientific reports

Radiomics: the facts and the challenges of image analysis