Using Reactive Machine AI to Derive Quantitative Lung CT Metrics of COPD, ILD, and COVID-19 Pneumonia





Introduction


This chapter will first review the basic structure of the human lung and use this information to explore the different QCT lung metrics that can be obtained from a single chest CT scan obtained at total lung capacity (TLC). A chest CT scan obtained at TLC and analyzed using lung CT AI can detect and assess lung density changes that occur in patients with emphysema from COPD, pulmonary inflammation and fibrosis in ILD, and acute viral pneumonia from COVID-19. The changes in lung density that result from these different diseases reflect important structural changes in the lung tissue that correlate with other measures of lung disease, such as clinical symptoms, exercise limitations, and pulmonary function testing. The successful application of lung CT AI to the assessment of diffuse lung diseases depends on the following four important steps: (1) quantitative chest CT protocol to acquire high-quality 3D CT images of both lungs, (2) segment the lungs from the rest of the thoracic anatomy, (3) extract quantitative CT metrics from the lung CT images and, (4) use the extracted QCT metrics to detect and assess normal and diseased lung tissue ( Fig. 5.1 ).




Fig. 5.1


Flow diagram outlining the important steps in a Reactive AI algorithm to assess a TLC lung CT scan for evidence of normal and diseased lung tissue.


Normal Lung Structure


The human lung is the largest visceral organ in the human body with a volume between 4 and 6 liters in normal adults. The lung is comprised mainly of air and water. The lung has high intrinsic contrast for x-ray CT imaging because the lung is 80% air and 20% water with HU values of −1000 HU and 0 HU, respectively. The high intrinsic contrast in the lung between water density and air density enables high-quality CT images of normal lung tissue, airways, and blood vessels.


There are 23 generations of airways from the trachea to the alveoli. The trachea is the first airway generation. There are two functional compartments of the lung airways: conducting airways, airway generations 1 to 16, airway generations 17 to 23, and gas diffusion ( Fig. 5.2 ). The conducting airways transport air from the largest conducting airway, the trachea, to the smallest conducting airway the terminal bronchiole, generation 16. The terminal bronchial conducts air to the lung acinus, which is the largest gas exchanging unit of the lung. The structure of the acinus includes several generations of respiratory bronchioles, alveolar ducts, and alveoli ( Fig. 5.3 ). The lung is designed to provide a very efficient transfer of oxygen and carbon dioxide gases. The oxygen molecules in inspired air are transferred from the alveolar spaces to the red blood cells, and carbon dioxide is transferred from red blood cells to the alveolar air spaces. The essential structural features of the lung include the high surface to volume ratio of the lung structure, as well as the thin alveolar walls that enable a very efficient exchange of oxygen from the alveolar lumen into the capillary lumen within the wall of the alveolus, and the efficient exchange of carbon dioxide from the red blood cells in the capillary lumen into the alveolar space. The surface area of the alveolar walls in a human lung is the size of a tennis court, 140 m 2 , but are folded into a very compact space 6 liters in size. The normal thickness of the alveolar wall is about 2 microns.




Fig. 5.2


Different generations of conducting and diffusion airways of the human lung.



Fig. 5.3


Terminal bronchiole, respiratory bronchioles, alveolar ducts, and alveoli that make up the pulmonary acinus of the lung as described in the text. The pulmonary acinus is the largest gas exchanging unit of the human lung.


QCT Scanning Protocol and Lung Segmentation


The first step in lung CT AI of diffuse lung disease is to obtain a quality 3D chest CT scan using an appropriate QCT scanning protocol that we described in detail in Chapter 3 . In this chapter, we will discuss using a single TLC CT scan to assess normal and diseased lung structure. In Chapter 6 , we will discuss how to obtain functional lung information by using both a TLC chest CT scan and an FRC/RV chest CT scan.


The second step is to automatically and consistently segment the lungs from the rest of the chest anatomy using a validated software program for this purpose. The computer algorithms that do this are quite sophisticated and use reactive AI, limited memory AI, or a combination of these AI levels. These AI algorithms are designed to run automatically and are very efficient. Fig. 5.4 shows an axial, sagittal, and coronal chest CT image from a normal patient. Lower density is represented by darker gray colors and higher dense tissues with lighter gray colors. The low density of normal lung tissue reflects the fact that the normal lung is 80% air and 20% soft tissue (10% blood and 10% tissue). The solid cylindrical branching soft tissue density structures in the lung are the arteries and veins. The hollow cylindrical branching air-containing structures are the airways. Fig. 5.5 shows a 3D lung CT image after the image segmentation software has processed the original chest CT images.




Fig. 5.4


(A) axial, (B) coronal, and (C) sagittal 2D planar images of the thorax at the level of the carina at the bifurcation of the trachea (arrows) . These images were displayed using a WW of 1500 HU and a WL of −500 HU. These WW and WL settings optimize the chest CT images for displaying the lung tissue for visual assessment.



Fig. 5.5


3D semitransparent surface rendering of the lungs with only the trachea and lung tissue remaining. The rest of the chest anatomy (e.g., chest wall, spine, heart, aorta) has been removed. The large central pulmonary artery and veins have also been removed. The lung voxel histograms that we discuss in this chapter are derived from the lung portrayed here with the central airways and central pulmonary vessels removed. The airways and central pulmonary vessels are usually assessed separately. The central airways are discussed toward the end of Chapter 6 and the central pulmonary vessels are discussed in Chapter 8.

(Courtesy of VIDA.)


The third step of lung CT AI to assess diffuse lung disease is to extract quantitative CT features from the segmented lung CT images. This can be used in step four to detect and assess normal and diseased lung tissue. In this chapter, we will discuss straightforward reactive lung CT AI methods to derive CT features from the lung voxel histogram ( Fig. 5.6 ). The fourth step of lung CT AI is to use the CT features derived in step three to assess the presence and extent of diffuse lung disease. Normal lung density, decreased lung density from emphysema, and increased lung density from pulmonary fibrosis and pneumonia can be assessed using Reactive Lung CT AI.




Fig. 5.6


TLC lung CT voxel histogram plot from a normal patient. The mean lung density is −860 HU in this normal patient.


Chronic Obstructive Pulmonary Disease (COPD) Induced Changes in Lung Structure


COPD first produces narrowing and destruction of small conducting airways before emphysema develops in the human lung. The narrowing and destruction of small conducting airways increase the resistance to airflow in COPD. The progression of COPD can then produce emphysema in the lung. “Emphysema is defined as a condition of the lung characterized by abnormal, permanent enlargement of airspaces distal to the terminal bronchiole, accompanied by the destruction of alveolar walls, and without obvious lung fibrosis”. This enlargement of alveoli and the destruction of alveolar walls effectively reduces the available surface area for gas exchange per unit volume of lung tissue; therefore the efficiency of gas exchange in the lung is decreased in emphysema. The destruction of tissue in emphysema will decrease the density of lung tissue, and this can be detected and assessed using lung CT AI.


Quantitative CT Metrics of Lung Density in COPD


3D CT images of the lungs can be analyzed by looking at the location and individual values of the lung CT voxels in the lung. The simplest approach is to assess the lung CT voxel histogram of both lungs. The spatial information is lost in this approach if the voxel histograms of both lungs are combined. Fig. 5.6 shows the voxel histogram plot of normal lungs. Fig. 5.7 shows the voxel histogram plot of emphysematous lungs.




Fig. 5.7


TLC lung CT voxel histogram from a patient with severe emphysema. The mean lung density is −890 HU and the LAA −950 is 39%. A vertical red line marks the −950 HU value.


Assessing features of the lung CT voxel histogram was one of the earliest methods of quantitatively assessing normal and diseased lung tissue. The disadvantage of combining the CT lung voxels from both lungs is that the spatial information is lost. This can be overcome by assessing the CT voxel histogram at smaller scales. This has been done for individual lungs, lung lobes, sublobar segments, and individual voxels. The voxel-level completely preserves the spatial and CT voxel value information for each voxel, and can then be further processed with more powerful lung CT AI methods; more on this in Chapter 6 . The following discussion will elaborate on different quantitative lung metrics that can be derived from the lung CT voxel histogram from a single TLC lung CT scan, and how they can be used to detect and assess normal and diseased lung tissue.


Mean Lung Density (MLD) for the Detection and Assessment of Emphysema


There are several ways to analyze the lung CT voxel histogram curve ( Figs. 5.6 and 5.7 ) that can provide meaningful quantitative CT metrics of normal lung tissue and lung tissue with emphysema. Emphysema decreases the density of the lung tissue due to the destruction of alveolar walls and capillaries within those walls and also increases the volume of the lungs due to decreased elastic recoil of the lung tissue ( Fig. 5.8 ). The decrease in lung tissue density in the emphysematous lung tissue, compared to normal lung tissue, can be detected by assessing the values of the lung CT voxels. The mean lung density (MLD) can be assessed by calculating the mean value of the lung CT voxels, and the CT determined MLD will decrease if there is emphysema in the lung. MLD density in normal subjects is greater than the MLD in patients with emphysema; an early research study reported this in 1982. The histogram curves from a normal lung and a lung with a lot of emphysema are shown in Figs. 5.6 and 5.7 . The MLD is indicated in these figures and you can see that the emphysematous lung has a mean lung density that is lower than the mean MLD of a normal lung. You can also appreciate that the shape of the histogram has also changed. Mean lung density, standard deviation, skewness, and kurtosis of the lung CT voxel histogram curve will change between normal lung and emphysematous lung, and also between normal lung and pulmonary fibrosis from ILD. This will be discussed further in this chapter.




Fig. 5.8


Two axial chest CT images. ( A) Normal lung with normal lung density (arrows) . ( B) Focal areas of decreased density (arrows) that are due to tissue destruction from emphysema.


Low Attenuating Area (LAA) for the Detection and Assessment of Emphysema


Another approach to assessing the lung CT voxel histogram for the presence or absence of emphysema is to use a threshold approach where the number of voxels below a certain voxel value in HU is assessed. There are two popular approaches to assessing the number of lung CT voxels below a certain threshold. The first approach picks a fixed voxel value and assesses the number of lung CT voxels that fall below this voxel value or threshold. This is referred to as the density mask, or low attenuating area (LAA) method, and is usually expressed as a percentage of lung CT voxels that are less than the threshold. The currently accepted LAA threshold for severe emphysema is −950 HU and expressed as LAA −950 ( Fig. 5.7 ). The second approach picks a fixed fraction or percentage of low attenuating voxels (e.g., 15%) in the lung CT voxel histogram and then assesses the lung CT voxel value that separates the lowest 15% of lung voxel values from the remaining higher 85% of voxel values. The currently accepted percentile method is the 15th percentile method.


In 1988 Muller et al. reported their results using an LAA thresholding method they described as the “Density Mask” technique to detect and assess emphysema. The density mask software in this study measures the number of lung voxels less than a certain threshold. The study looked at threshold values of −900 HU, −910 HU, and −920 HU; the −910 HU threshold gave the best results in this study. There were 28 subjects in this study. The CT scans were obtained on a single row fan-beam detector array, third-generation axial CT scanner, GE 9800. The entire lung was scanned using 10-mm-thick axial images obtained 10 mm apart, contiguous images of the whole lungs. Each of the subjects was referred for resection of their lung cancers after their CT scans. The study assessed visual evidence of emphysema, quantitative CT evidence of emphysema looking at the amount of lung, >−910 HU, and pathologic evidence of emphysema on the resected lung specimens. The same area of lung was examined by visual CT, quantitative CT, and pathological assessment of the lung. There was excellent agreement between all three methods. Intravenous contrast material was administered to each subject and this may have raised the optimal cutoff point in this study.


One of the quantitative CT challenges for assessing emphysema in the 1990s was trying to determine the optimal threshold for the LAA method. There were two key papers published in 1995 and 1996 by Gevenois et al. that established the optimal LAA threshold voxel value was −950 HU. The LAA −950 that was established by Gevenois is still used today. The optimal LAA threshold is also a function of the voxel size and the reconstruction kernel. The studies by Gevenois used a voxel that was approximately 1 mm × 1 mm × 1 mm, whereas the study by Mueller, used a voxel size of approximately 10 mm × 10 mm × 10 mm. The smaller the voxel size the better small emphysematous lesions, <5 mm, can be detected and a lower LAA threshold value is justified.


Let us look more closely at the two studies by Gevenois that concluded the optimal LAA threshold for emphysema was −950 HU. These two important studies looked at the correlation between macroscopic and microscopic pathologic evidence of emphysema and the QCT LAA −950 metric. The first study, in 1995, showed an excellent correlation between the area of emphysema on macroscopic horizontally/axially sectioned lung pathology specimens and the QCT LAA −950 HU metric for emphysema on high-resolution CT images of the lungs, xyz-plane resolution <2 mm. The axial pathology specimens were macroscopic tissue sections obtained every 1 to 2 cm from the resected lobe(s) or lung. In this study, 1.25 mm axial CT images of the lungs were obtained at 1-cm intervals, and these were assessed for emphysema using the LAA method where the fraction of lung voxels, >−950 HU, is expressed as a percent (e.g., LAA −950 = 15%). This study looked at a number of thresholds from −900 HU to −970 HU. The LAA −950 HU level had the best agreement with the macroscopic scores for emphysema on the axial pathology specimens of the lungs.


The second paper, published in 1996, examined the lung tissue obtained in the 1995 paper using microscopic methods of assessing normal lung tissue and emphysematous lung tissue. This included assessing the mean perimeter (MP) of the alveoli and alveolar ducts of 35 randomly selected microscopic fields, 25× magnification, and the mean interwall distance (MIWD) of alveoli and alveolar ducts within the 35 randomly selected fields. The randomly selected microscopic sections were assessed by experienced pathologists to classify them into normal lung or emphysematous lung. The MP and MIWD values were correlated with the CT LAA values for a range of LAA thresholds from −900 HU to −970 HU by 10 HU increments. The strongest correlation between MP and MIWD was the LAA set at −950 HU. There were multiple strong correlations between the MP and MIWD values and measures of the patient’s pulmonary function tests, with the strongest correlations being between DLCO/VA percent predicted and MP, MIWD, and Macroscopic evidence of emphysema. Setting the LAA threshold to −950 HU is now generally accepted as the best threshold value to use in an LAA method for the detection and assessment of emphysema on a TLC chest CT scan.


15th Percentile Method for the Detection and Assessment of Emphysema


The 15th percentile method of assessing the lung CT voxel histogram for evidence of emphysema picks a fixed fraction or percentage of low attenuating voxels (e.g., 15%) of the lung CT voxel histogram, and then assesses the lung CT voxel value that separates the lowest 15% of lung voxel values from the remaining higher 85% of voxel values. The percentile method was first introduced by Gould et al. in 1988.


Gould et al. described in 1988 the close correlation of low-density areas seen on chest CT with both macroscopic and microscopic measures of emphysema. The microscopic assessment of emphysema included measuring the surface area of distal airspace walls (e.g., alveolar walls) to the unit lung volume ratio (AWUV) of the lung in patients with mild to moderate emphysema who had preoperative CT scans and pulmonary function testing and then subsequently underwent lobe or lung resection for nonsmall cell lung cancer, which is a frequent unfortunate complication in heavy smokers with and without COPD. There was a strong correlation in this study between the EMI CT number of the fifth percentile and microscopic evidence of emphysema using the mean AWUV. There was also a strong correlation between the mean AWUV and pulmonary function tests. These results established that assessing the lung CT voxel histogram using the percentile method is a valid noninvasive test to detect and assess emphysema in COPD patients. There are several points worth discussing further in regards to this early high impact paper. First, the CT scanner used was an EMI 5005 CT scanner that obtained a 1.5 mm × 1.5 mm × 13 mm, axial image of the thorax with a scanning time of 17 seconds. The authors indicated that each CT image was obtained within 500 mL of TLC, though it is not clear how this was determined. The EMI unit is half the value of the Hounsfield unit previously discussed in this section. In this study by Gould, they recommended viewing the lung tissue for emphysema by highlighting the voxels in the CT image between −500 EMI units and −450 EMI units. This can be expressed as LAA using EMI units rather than HU units (e.g., LAA −450EMI ). This would correspond to −1000 HU and −900 HU. This can be viewed as a LAA −900HU approach to assessing emphysema, as discussed in the LAA section on detection and assessment of emphysema. Gould did not assess the LAA −450EMI quantitatively. The fifth percentile method in this study would translate to the 10% method using HU units for the voxel histogram rather than the EMI units of the original paper. Gould’s 1988 paper found that the CT voxel fifth percentile method for the assessment of emphysema was highly correlated to the mean AWUV, R = −0.63. Gould also showed that the mean AWUV was highly correlated with the DLCO/Va, R = 0.66. The authors did not assess the correlation of the LAA −450EMI with mean AWUV. We do not know why LAA was not investigated. The fifth percentile number varied from −479 EMI to −414 EMI units or −958 HU to −828 HU. The authors conclude that the results of their study indicate that CT measures of lung density and DLCO/VA in the living patient can objectively assess the alveolar surface to volume ratio, and can also localize spatially those portions of the lung that are contributing to the loss of tissue and expansion of airspaces which the DLCO/VA cannot. This paper was written in 1988 and this book is being written in 2020; we are now at the point of using CT measures of lung density as an objective way to assess the decreasing alveolar surface to volume ratio that occurs in COPD subjects with emphysema in routine clinical care of patients. The path was clear several decades ago, but many technological advances in lung CT AI were necessary before the automatic assessment of the lung CT voxel histogram could be integrated into routine clinical care, see Chapter 9 .


The 15th percentile method was used by Dirksen et al. in the late 1990s to build on the work of Gould, that we have just described. Dirksen’s work used improvements in CT scanner technology to obtain CT scans of the entire lung in a single breath-hold. This was made possible by using the spiral CT scanning mode that was not available to Gould in the late 1980s.


The 15th percentile method determines the CT number in HU that determines the point in the lung voxel histogram curve where 15% of the lung voxels are less than this number and 85% are greater than this number. The idea here is that as emphysematous lung tissue replaces normal lung tissue, the value of the Perc15 HU value will shift toward lower values. There was a detailed study looking at LAA cutoffs and percent cutoffs by Dirksen; they showed the optimal percentile for the percentile method was the 12th percentile and that the 12th percentile method performed better than LAA using level −930 HU in assessing longitudinal changes in the amount of emphysema present in subjects with alpha-1 antitrypsin deficiency (A1AD). This study looked at the percentiles from 1% to 50% by 1% increments and found the best results were in the range of 10% to 20%. A subsequent study (see Chapter 3 ) by Dirksen looking at patients with A1AD for a period of 3 years used the 15th percentile method to detect and assess A1AD related emphysematous lung tissue. Dirksen reported in 1999 that quantitative CT using the 15th percentile point of the lung density histogram showed decreased progression of A1AD induced emphysema in A1AD patients who were receiving monthly intravenous augmentation therapy with human alpha-AT. The quantitative CT results were much more sensitive than monitoring decreases in airflow measured using the forced expiratory volume in one second, FEV1. This was a turning point in the development of lung CT AI using quantitative CT metrics, such as the 15th percentile method, because now quantitative lung CT was doing something noninvasively that visual lung CT and pulmonary physiological measures of lung function could not do.


Clinical Value of Using Lung CT AI in Patients with Environmental Exposure to Cigarette Smoke


Both LAA −950 and the 15th percentile methods are two quantitative CT metrics that were both developed some time ago and persist today as the two favored methods of assessing the emphysema-related structural changes in the lung that reduce lung density. The LAA −950 and the 15th percentile are derived from the lung voxel histogram curves from a single chest CT scan obtained at TLC. There has been considerable research done looking at patients with COPD and determining what the clinical benefit of lung CT AI is in COPD patients.


Clinical Benefit of LAA −950


The COPD Genetic Epidemiology study (COPDGene) is a large, 10,263 subjects enrolled in phase 1, ongoing multicenter NIH-funded research study that began in 2007. Recently the study has reported results using the LAA −950 method of the lung CT voxel histogram to detect and assess emphysema at baseline and 5 years. In the phase 1 study, 8784 subjects were evaluated to assess subsequent mortality risk. Five years later, 4925 returning subjects who were studied at baseline in the phase 1 study were evaluated to assess the risk of decline in FEV1. The current method of diagnosing COPD requires postbronchodilator pulmonary function testing in a patient suspected of having COPD by using Spirometry to assess airflow, FEV1, lung volume, FVC, and the ratio of FEV1/FVC to make the diagnosis of COPD. The diagnosis of COPD requires that the FEV1/FVC ratio is less than 0.70.


A substantial number of patients with COPD manifest respiratory symptoms and CT structural evidence of COPD, suggesting COPD, before they develop an FEV1/FVC ratio less than 0.70. The COPDGene study has recently shown that these same patients are at increased risk of death and progressive decline in their FEV1/FVC ratio, and FEV1 percent predicted. The COPDGene study used the following four novel criteria to diagnose COPD and to assess the impact of these criteria on COPD progression and mortality: “Exposure, Symptoms, CT Structural Abnormality, Spirometry” ( Box 5.1 ). Exposure was defined as having a 10-pack-year or greater smoking history. Symptoms were defined as self-reporting a modified Medical Council (mMRC) dyspnea score of 2 or greater, and/ or chronic bronchitis (self-reported chronic cough and phlegm). CT structural abnormality was defined as having one or more of the following: LAA −950 equal to 5% or greater on a TLC chest CT scan (QCT structural measure of Emphysema), LAA −856 equal to or greater than 15% on expiratory CT scan (QCT functional measure of air trapping), Pi10 equal to 2.5 mm or greater (QCT structural measure of airway wall thickening). Spirometry evidence of COPD was defined as having an FEV1/FVC less than 0.70 and/or an FEV1 percentage predicted less than 80%. This COPDGene study proposed a scheme of assessing subjects with respiratory symptoms suspicious of COPD by placing them into one of eight categories, A thru H, using the four novel criteria just described: Exposure, Symptoms, CT Structural Abnormalities, Spirometry ( Box 5.2 ). Category A is a patient with significant exposure to cigarette smoke. Category B is a patient with significant exposure to cigarette smoke and QCT evidence of COPD. Category C is a patient with significant exposure to cigarette smoke and clinical symptoms of COPD. Category D is a patient with significant exposure to cigarette smoke and abnormal spirometry with FEV1/FVC ratio less than 0.70 and/or FEV1 percent predicted less than 80%. Category E is a patient with significant exposure to cigarette smoke, QCT evidence of COPD, and clinical symptoms of COPD. Category F, is a patient with significant exposure to cigarette smoke, clinical symptoms of COPD, and abnormal spirometry with FEV1/FVC ratio less than 0.70 and/or FEV1 percent predicted less than 80%. Category G, is a patient with significant exposure to cigarette smoke, QCT evidence of COPD, and abnormal spirometry with FEV1/FVC ratio less than 0.70 and/or FEV1 percent predicted less than 80%. Category H is a patient with significant exposure to cigarette smoke, QCT evidence of COPD, clinical symptoms of COPD, and abnormal spirometry with FEV1/FVC ratio less than 0.70 and/or FEV1 percent predicted less than 80%. Categories A thru H are summarized in Box 5.2 .


Feb 6, 2023 | Posted by in CARDIOLOGY | Comments Off on Using Reactive Machine AI to Derive Quantitative Lung CT Metrics of COPD, ILD, and COVID-19 Pneumonia

Full access? Get Clinical Tree

Get Clinical Tree app for offline access