Quantitative CT Analysis of Diffuse Lung Disease

Published Online:https://doi.org/10.1148/rg.2020190099

Abstract

Quantitative analysis of thin-section CT of the chest has a growing role in the clinical evaluation and management of diffuse lung diseases. This heterogeneous group includes diseases with markedly different prognoses and treatment options. Quantitative tools can assist in both accurate diagnosis and longitudinal management by improving characterization and quantification of disease and increasing the reproducibility of disease severity assessment. Furthermore, a quantitative index of disease severity may serve as a useful tool or surrogate endpoint in evaluating treatment efficacy. The authors explore the role of quantitative imaging tools in the evaluation and management of diffuse lung diseases. Lung parenchymal features can be classified with threshold, histogram, morphologic, and texture-analysis–based methods. Quantitative CT analysis has been applied in obstructive, infiltrative, and restrictive pulmonary diseases including emphysema, cystic fibrosis, asthma, idiopathic pulmonary fibrosis, hypersensitivity pneumonitis, connective tissue–related interstitial lung disease, and combined pulmonary fibrosis and emphysema. Some challenges limiting the development and practical application of current quantitative analysis tools include the quality of training data, lack of standard criteria to validate the accuracy of the results, and lack of real-world assessments of the impact on outcomes. Artifacts such as patient motion or metallic beam hardening, variation in inspiratory effort, differences in image acquisition and reconstruction techniques, or inaccurate preprocessing steps such as segmentation of anatomic structures may lead to inaccurate classification. Despite these challenges, as new techniques emerge, quantitative analysis is developing into a viable tool to supplement the traditional visual assessment of diffuse lung diseases and to provide decision support regarding diagnosis, prognosis, and longitudinal evaluation of disease.

©RSNA, 2019

SA-CME LEARNING OBJECTIVES

After completing this journal-based SA-CME activity, participants will be able to:

  • ■ Discuss the basic statistics and machine learning principles involved in quantitative CT analysis.

  • ■ Describe various applications of quantitative analysis in diffuse lung disease.

  • ■ Recognize potential pitfalls in interpreting automated quantitative output.

Introduction

Despite sharing similar symptoms and clinical features, diffuse lung diseases are a heterogeneous group of pathologically distinct processes that have markedly different prognoses and treatment options (1). In addition to pulmonary function tests (PFTs), thin-section CT of the chest plays a critical role in distinguishing different diseases, evaluating disease severity, following disease progression, and monitoring treatment response. At thin-section CT, the extent and distribution of specific imaging characteristics such as honeycombing, ground-glass opacities, traction bronchiectasis, reticulation, and mosaic attenuation can be suggestive of or pathognomonic for specific diagnoses. For instance, cases demonstrating characteristic thin-section CT features of usual interstitial pneumonia (UIP) do not require biopsy confirmation for definitive diagnosis and therapeutic determination (2). Extensive ground-glass opacities or areas of air trapping or mosaic attenuation may suggest an alternative diagnosis. However,

diagnostic certainty in the absence of characteristic UIP features is challenging owing to the substantial overlap of imaging phenotypes, heterogeneity of presentation within a disease, and subjective differences in detection or interpretation of findings (35).
Quantitative tools can assist in both accurate diagnosis and longitudinal management by improving characterization through augmented displays of disease distribution and quantification of disease in an objective and reproducible manner.

There are limitations to the traditional methods of using physiologic data such as PFTs and visual thin-section CT findings in concert to grade interstitial lung disease (ILD) severity and guide therapy. Commonly used pulmonary function measurements quantify the degree of physiologic impairment but can be relatively insensitive to marginal changes or early disease and do not reveal the morphologic changes or spatial distribution of abnormalities that can help identify the cause (6,7). Inter- and intraobserver variations in visual assessment of thin-section CT findings increase the challenge of disease assessment (3,8). Quantitative imaging may play an important role in this respect by increasing the reproducibility of evaluation and producing potential tools that can serve to monitor disease progression and treatment response. This may be particularly relevant in situations of subtle change as new therapies are developed and evaluated (6). For instance, idiopathic pulmonary fibrosis (IPF) treatment may involve the administration of relatively new medications such as pirfenidone, nintedanib, PRM-151, or other disease-modifying treatments (911).

A number of quantitative analysis techniques have been developed to evaluate diffuse lung disease. The purpose of this article is to review the basic concepts of quantitative CT analysis and model development, examples of clinical applications in diffuse lung disease, challenges in analysis and the use of quantitative data as decision support in interpretation, and future directions for research.

Basic Concepts of Quantitative CT Analysis

Quantitative CT analysis is one area in the rapidly growing field of radiomics, the practice of extracting, analyzing, and interpreting quantitative data from medical images to aid in disease diagnosis and prognosis (12) (Table). Most quantitative tools use volumetric datasets. Different types of information can be extracted from the data.

Quantitative analysis can range from simple threshold measurements that count individual pixels within a range of values to texture metrics that capture morphology and regional heterogeneity or complex spatiotemporal algorithms that can be used to determine changes in features over time.

Terminology Used in Quantitative CT Analysis of Diffuse Lung Disease

Threshold-based Measures

Threshold-based measures generally count pixels above or below a certain attenuation value and sum these counts to determine an absolute or relative volume of these features. For example, the extent of areas of abnormally low attenuation, as measured by the percentage of voxels below the attenuation of normal lung parenchyma, is useful in approximating the extent and severity of diseases that result in lower lung density, such as emphysema. However, threshold-based measures that do not assess morphology or local spatial relationships cannot differentiate between different diseases with similar attenuation values that are morphologically or pathologically distinct. For example, these measures have limited utility in distinguishing paraseptal emphysema from honeycombing, distinguishing emphysema from air trapping and cystic lesions, and differentiating reticular fibrosis from ground-glass opacities and consolidations in ILD (13).

First-Order Statistics

Threshold measurements are typically not used directly in texture analysis. Instead, statistical calculations such as first-order statistics, including histogram-based summary statistics for the whole lung or a region of interest (ROI), are employed. The mean, median, histogram skew, and kurtosis of attenuation can reflect the extent or severity of diffuse processes, such as emphysema or pulmonary fibrosis. For example, alveolar destruction in emphysema leads to abnormally lucent parenchyma resulting in decreased mean attenuation. In comparison, fibrotic changes in a purely fibrotic process are higher in attenuation compared with that of normal parenchyma, engendering an increase in mean lung attenuation.

However, first-order features do not take into account the spatial relationship between voxels. Whole-lung statistics without this spatial information can be confusing in mixed obstructive and restrictive processes.

Higher-Order Textural Analysis

Second-order and higher statistics take into account the spatial relationship between ROIs or voxels and characterize textural features. Methods such as run-length matrices, fractal measures, and gray-level co-occurrence matrices can be used to determine ROI uniformity, shape, and other morphologically distinct features. The unique thin-section CT patterns of diffuse lung disease such as honeycombing or traction bronchiectasis are well suited for textural analysis. Additionally, different types of image filtration can be performed to remove noise, enhance edges, and emphasize or extract certain features (14). Ultimately, several of these textural features can be used together in an algorithm to help create a detection and prediction model (1518).

Quantitative analysis tools produce objective, quantifiable, and reproducible assessments of pulmonary parenchymal changes associated with diffuse lung diseases.

Model Development and Machine Learning

Several quantitative texture analysis tools (15,1923), including DTA and CALIPER, use machine learning. Broadly, machine learning is a branch of artificial intelligence in which an algorithm, when presented with a set of data, can distill the salient features of a training dataset through complex computational statistics and mathematical optimization methods. Then the mathematical model and the features discovered in the training dataset can be used to make predictions and classify future input (24). The machine or model learns in that it algorithmically optimizes its performance to reach the correct answer on the basis of the input features with new data input.

There are two broad types of machine learning: supervised and unsupervised. In supervised learning, the machine is given prelabeled data, including the number and type of groups. In unsupervised learning, the machine is given data without any prelabeling and must independently determine how to group the data. To develop a supervised machine learning algorithm to classify radiologic images, the machine is initially given a set of labeled training data from which it then develops a method for predicting the labels while continuing to improve with new input. Then, the machine is given an unlabeled validation and/or test data on which it can apply its learned model. For example, to create a supervised machine learning algorithm that can distinguish among diffuse lung diseases, a set of thin-section CT images labeled with the diagnoses is initially provided for learning purposes, followed by a set of unlabeled validation thin-section CT images on which the algorithm can be tested. In comparison, in unsupervised learning, the same set of thin-section CT images is given without revealing or labeling the diagnoses.

While some of the machine learned quantitative tools are primarily based on supervised learning, others use a combination of both supervised and unsupervised learning. For example, DTA used unsupervised clustering analysis on randomly sampled parenchyma from CT images of patients with IPF and nonsmoking controls to produce a dictionary of low-level features that distinguish fibrosis from nonfibrotic lungs (19). Then, separately, radiologist-labeled ROIs were created from the same cohort of IPF patients. Quantitative features for these ROIs were then computed by matching these ROIs with the previously constructed dictionary of low-level features. The resultant labeled exemplars were then used to train the supervised support vector machine classifier to distinguish fibrosis from normal lungs. The DTA fibrosis score is the number of ROIs classified as fibrosis divided by the total number of ROIs sampled from the lung segmentation volume.

Similar to DTA, CALIPER, another textural analysis tool for diffuse lung diseases, also utilizes a combination of supervised and unsupervised learning. However, CALIPER begins with supervised learning. Thin-section CT images of 14 patients with pathologically proven ILD from the Lung Tissue Research Consortium database were used, and the lung parenchyma was divided into volumes of interest (VOIs) (25). Randomly selected VOIs were displayed and categorized by four expert thoracic radiologists into one of five lung parenchymal patterns: normal, ground-glass opacities, reticular infiltrate, honeycombing, and low-attenuation areas and/or emphysema (Fig 1). Next, different metrics based on first- and second-order histogram-based statistical descriptors were tested for their ability to discriminate among the five categories, and a multidimensional scaling representation of the Cramér–von Mises distance was chosen. The training VOIs were then automatically clustered and stratified into unbiased groups using affinity propagation, a type of unsupervised machine learning algorithm, based on their Cramér–von Mises values. The VOI at the center of each cluster was determined to be the key exemplar for that feature to be used in the future classification of input data. Prospectively, when a given thin-section CT image needs to be characterized, each voxel is classified on the basis of analysis of its local surrounding VOI histogram features (Fig 2a). This process is repeated until both lungs are completely characterized. The findings are then summarized for the entire lung and can be used for statistical analysis with absolute volumes or percentage of each feature per region or visually represented as a color overlay over the original cross-sectional images or as a summary glyph (Fig 2b).

Figure 1a.

Figure 1a. Expert-identified parenchymal pattern exemplars. (a–e) Axial thin-section CT images show patterns of lung parenchyma classified by CALIPER as normal parenchyma (a), ground-glass opacities (arrows in b) (b), reticular densities (arrows in c) (c), honeycombing (arrows in d) (d), and low-attenuation areas (arrows in e) (e). (f) Diagram shows CALIPER 15 × 15 × 15 voxel VOIs (cubes) that match the parenchymal patterns, with conical signatures and associated color codes (rectangles) assigned to each pixel in a dataset.

Figure 1b.

Figure 1b. Expert-identified parenchymal pattern exemplars. (a–e) Axial thin-section CT images show patterns of lung parenchyma classified by CALIPER as normal parenchyma (a), ground-glass opacities (arrows in b) (b), reticular densities (arrows in c) (c), honeycombing (arrows in d) (d), and low-attenuation areas (arrows in e) (e). (f) Diagram shows CALIPER 15 × 15 × 15 voxel VOIs (cubes) that match the parenchymal patterns, with conical signatures and associated color codes (rectangles) assigned to each pixel in a dataset.

Figure 1c.

Figure 1c. Expert-identified parenchymal pattern exemplars. (a–e) Axial thin-section CT images show patterns of lung parenchyma classified by CALIPER as normal parenchyma (a), ground-glass opacities (arrows in b) (b), reticular densities (arrows in c) (c), honeycombing (arrows in d) (d), and low-attenuation areas (arrows in e) (e). (f) Diagram shows CALIPER 15 × 15 × 15 voxel VOIs (cubes) that match the parenchymal patterns, with conical signatures and associated color codes (rectangles) assigned to each pixel in a dataset.

Figure 1d.

Figure 1d. Expert-identified parenchymal pattern exemplars. (a–e) Axial thin-section CT images show patterns of lung parenchyma classified by CALIPER as normal parenchyma (a), ground-glass opacities (arrows in b) (b), reticular densities (arrows in c) (c), honeycombing (arrows in d) (d), and low-attenuation areas (arrows in e) (e). (f) Diagram shows CALIPER 15 × 15 × 15 voxel VOIs (cubes) that match the parenchymal patterns, with conical signatures and associated color codes (rectangles) assigned to each pixel in a dataset.

Figure 1e.

Figure 1e. Expert-identified parenchymal pattern exemplars. (a–e) Axial thin-section CT images show patterns of lung parenchyma classified by CALIPER as normal parenchyma (a), ground-glass opacities (arrows in b) (b), reticular densities (arrows in c) (c), honeycombing (arrows in d) (d), and low-attenuation areas (arrows in e) (e). (f) Diagram shows CALIPER 15 × 15 × 15 voxel VOIs (cubes) that match the parenchymal patterns, with conical signatures and associated color codes (rectangles) assigned to each pixel in a dataset.

Figure 1f.

Figure 1f. Expert-identified parenchymal pattern exemplars. (a–e) Axial thin-section CT images show patterns of lung parenchyma classified by CALIPER as normal parenchyma (a), ground-glass opacities (arrows in b) (b), reticular densities (arrows in c) (c), honeycombing (arrows in d) (d), and low-attenuation areas (arrows in e) (e). (f) Diagram shows CALIPER 15 × 15 × 15 voxel VOIs (cubes) that match the parenchymal patterns, with conical signatures and associated color codes (rectangles) assigned to each pixel in a dataset.

Figure 2a.

Figure 2a. Texture-based analysis parenchymal classification summary. (a) Flowchart shows the process that CALIPER uses for characterizing lung parenchyma. For each 15 × 15 × 15 voxel-sized VOI, histogram-based characteristics are computed and then compared with the histograms of the exemplars of each parenchymal pattern using Cramér–von Mises dissimilarity measurements. The exemplar histogram with the least difference compared with the local histogram is assigned as the ultimate class of the underlying voxel. This process is repeated for each voxel until both lungs are classified. (b) Diagram shows texture analysis outputs summarized by glyphs. The entire glyph represents total lung volume and is divided by lung anatomy: right upper lung (RU), right middle lung (RM), right lower lung (RL), left upper lung (LU), left mid lung (LM), and left lower lung (LL). Each concentric circle represents 20% of the glyph. Keys are the same for Figures 4c, 4f, 5c, and 5f. The distribution and pattern of pulmonary parenchymal abnormalities may suggest a specific disease process, such as the bibasilar honeycombing of UIP (left), upper lung–predominant honeycombing of chronic hypersensitivity pneumonitis (CHP) (middle), or upper lung–predominant low-attenuation areas (LAA) of centrilobular emphysema (right).

Figure 2b.

Figure 2b. Texture-based analysis parenchymal classification summary. (a) Flowchart shows the process that CALIPER uses for characterizing lung parenchyma. For each 15 × 15 × 15 voxel-sized VOI, histogram-based characteristics are computed and then compared with the histograms of the exemplars of each parenchymal pattern using Cramér–von Mises dissimilarity measurements. The exemplar histogram with the least difference compared with the local histogram is assigned as the ultimate class of the underlying voxel. This process is repeated for each voxel until both lungs are classified. (b) Diagram shows texture analysis outputs summarized by glyphs. The entire glyph represents total lung volume and is divided by lung anatomy: right upper lung (RU), right middle lung (RM), right lower lung (RL), left upper lung (LU), left mid lung (LM), and left lower lung (LL). Each concentric circle represents 20% of the glyph. Keys are the same for Figures 4c, 4f, 5c, and 5f. The distribution and pattern of pulmonary parenchymal abnormalities may suggest a specific disease process, such as the bibasilar honeycombing of UIP (left), upper lung–predominant honeycombing of chronic hypersensitivity pneumonitis (CHP) (middle), or upper lung–predominant low-attenuation areas (LAA) of centrilobular emphysema (right).

Clinical Applications

Quantitative tools have been evaluated in diverse pulmonary diseases including but not limited to chronic obstructive pulmonary disease (COPD), IPF, hypersensitivity pneumonitis, connective tissue–related ILD, lymphangiomyomatosis (LAM), and combined pulmonary fibrosis and emphysema.

Chronic Obstructive Pulmonary Disease

The thin-section CT appearance of emphysema is characterized by an increase in low-attenuation areas, reflecting pathologic alveolar destruction. This is directly reflected in decreased lung density and is commonly quantified with threshold-based measures. One method calculates the percentage of total or lobar lung parenchyma below a certain attenuation threshold, commonly around –950 HU, generally considered to represent the extent of emphysema and often called the emphysema index or relative area less than –950 HU (RA –950 HU). A higher percentage reflects more extensive disease.

Another method, the 15th percentile technique (Perc15), is whole-lung histogram-based. On a histogram representing the attenuation of all lung voxels, this index represents the threshold value in Hounsfield units for which 15% of all lung voxels have a lower attenuation value. A lower 15th percentile value represents more severe disease. A higher quantitatively calculated baseline level of emphysema was found to be significantly related to lower baseline pulmonary function, greater decline in pulmonary function, and a higher rate of developing airway obstruction (26). These indices can be further examined on a lobar or subsegmental level. Perc15 has also been used to quantify the rate of lung tissue loss in α-1 antitrypsin deficiency (27,28) and COPD (29). In patients with concomitant fibrosis, the emphysema index is less likely to underdiagnose emphysema than a measure such as mean lung density, in which abnormal low and high densities essentially average each other out (Fig 3) (30).

Figure 3a.

Figure 3a. Threshold-based CT quantification. Emphysema can be quantified at CT based on threshold measurements, such as the volume or percentage of lung parenchyma with an attenuation of less than –950 HU. Axial (a, d) and coronal (b, c, e, f) CT images with color overlay (light blue = areas < –950 HU) and three-dimensional projections (c, f) in a patient with emphysema (a–c) and a disease-free control (d–f) show the characteristic upper lung–predominant pattern of centrilobular emphysema. Green outline = right lung, tubular red structures in c and f = airways, red outline = left lung.

Figure 3b.

Figure 3b. Threshold-based CT quantification. Emphysema can be quantified at CT based on threshold measurements, such as the volume or percentage of lung parenchyma with an attenuation of less than –950 HU. Axial (a, d) and coronal (b, c, e, f) CT images with color overlay (light blue = areas < –950 HU) and three-dimensional projections (c, f) in a patient with emphysema (a–c) and a disease-free control (d–f) show the characteristic upper lung–predominant pattern of centrilobular emphysema. Green outline = right lung, tubular red structures in c and f = airways, red outline = left lung.

Figure 3c.

Figure 3c. Threshold-based CT quantification. Emphysema can be quantified at CT based on threshold measurements, such as the volume or percentage of lung parenchyma with an attenuation of less than –950 HU. Axial (a, d) and coronal (b, c, e, f) CT images with color overlay (light blue = areas < –950 HU) and three-dimensional projections (c, f) in a patient with emphysema (a–c) and a disease-free control (d–f) show the characteristic upper lung–predominant pattern of centrilobular emphysema. Green outline = right lung, tubular red structures in c and f = airways, red outline = left lung.

Figure 3d.

Figure 3d. Threshold-based CT quantification. Emphysema can be quantified at CT based on threshold measurements, such as the volume or percentage of lung parenchyma with an attenuation of less than –950 HU. Axial (a, d) and coronal (b, c, e, f) CT images with color overlay (light blue = areas < –950 HU) and three-dimensional projections (c, f) in a patient with emphysema (a–c) and a disease-free control (d–f) show the characteristic upper lung–predominant pattern of centrilobular emphysema. Green outline = right lung, tubular red structures in c and f = airways, red outline = left lung.

Figure 3e.

Figure 3e. Threshold-based CT quantification. Emphysema can be quantified at CT based on threshold measurements, such as the volume or percentage of lung parenchyma with an attenuation of less than –950 HU. Axial (a, d) and coronal (b, c, e, f) CT images with color overlay (light blue = areas < –950 HU) and three-dimensional projections (c, f) in a patient with emphysema (a–c) and a disease-free control (d–f) show the characteristic upper lung–predominant pattern of centrilobular emphysema. Green outline = right lung, tubular red structures in c and f = airways, red outline = left lung.

Figure 3f.

Figure 3f. Threshold-based CT quantification. Emphysema can be quantified at CT based on threshold measurements, such as the volume or percentage of lung parenchyma with an attenuation of less than –950 HU. Axial (a, d) and coronal (b, c, e, f) CT images with color overlay (light blue = areas < –950 HU) and three-dimensional projections (c, f) in a patient with emphysema (a–c) and a disease-free control (d–f) show the characteristic upper lung–predominant pattern of centrilobular emphysema. Green outline = right lung, tubular red structures in c and f = airways, red outline = left lung.

Threshold-based methods can also be used to quantify the degree of air trapping. Air trapping can be quantified as the percentage of total lung pixels with an attenuation of –856 HU or less at expiratory CT. The threshold value of –856 HU was chosen because it is the expected attenuation of normal lung at inspiratory CT. Areas of normal lung will appropriately increase in attenuation on expiratory images, whereas areas of air trapping will not. This quantitative air-trapping measure has been shown to better correlate with spirometry findings than does the emphysema index (the percentage of lung pixels < –950 HU) in patients enrolled in the Genetic Epidemiology of COPD (COPDGene) Study (31). Similarly, among patients with asthma, it was able to identify a subset with more severe disease (32).

It can be challenging to distinguish between low-attenuation areas due to air trapping, which is secondary to small airway obstruction, and low-attenuation areas related to emphysema, which is due to alveolar destruction; the two often coexist in this patient population. One method to address this is through parametric response mapping (13), which uses a combination of threshold-based measures at coregistered inspiratory and expiratory CT. At expiratory imaging, areas of emphysema or air trapping can both be distinguished from normal lung as abnormally lucent areas that can be quantitatively identified by the air-trapping threshold measure of an attenuation less than or equal to –856 HU. On the coregistered inspiratory images, these abnormal areas can be further parsed by using the emphysema threshold measure of an attenuation less than or equal to –950 HU. Areas of air trapping should have normal attenuation on inspiratory images, whereas areas of emphysema will be abnormally lucent. The addition of parametric response mapping–derived measures to predictive models of COPD has increased their discriminative value.

There are variables that impact these threshold-based and first-order statistics that are problematic in the real-world setting of variable CT acquisition and reconstruction techniques. Specifically, although these methods have been shown to correlate with physiology and disease severity, the threshold measures are subject to a great deal of variation from technical factors such as volume averaging, section thickness, image noise variations owing to different radiation dose or patient size, and CT image reconstruction effects such as edge enhancement or smoothing techniques (33,34). In the COPDGene project, the correlation of visual versus semiautomated quantitative CT assessments showed moderate interobserver agreement for presence or absence of emphysema but lower levels of agreement for distinct emphysema subtypes (5). Threshold-based measures of emphysema severity were found to be only weakly related to lung cancer risk in a high-risk screening population (35). The correlation of quantitative CT features with pathologic emphysema or genotype is improved by combining multiple different types of measures. The relative area of lung attenuation below –960 HU, the relative area below –970 HU, and the first percentile values were significantly correlated with pathologic references of disease and, along with a ratio of forced expiratory volume in 1 second and vital capacity and diffusing capacity of the lung for carbon monoxide, were complementary in predicting the extent of microscopic disease (36,37).

Higher-order texture-based analysis may avoid some of the pitfalls of threshold measures. Textural tools are adept at characterizing the regional distribution and morphology of low-attenuation changes by segmenting the lung into small VOIs. This can be used to suggest specific subtypes of emphysema such as paraseptal or centrilobular or to differentiate among visually and pathologically distinct causes of diffuse low-attenuation areas, such as air trapping, cystic lung disease, and honeycombing (Fig 4).

Figure 4a.

Figure 4a. Texture-based CT quantification. Coronal thin-section CT images (a, b, d, e) and glyphs (c, f) show the various patterns of emphysema (arrows in a, b, d, e), such as upper lobe–predominant centrilobular emphysema (a–c) and lower lobe–predominant pan-lobular emphysema (d–f), without (a, d) and with (b, e) texture-based quantification overlay and corresponding glyphs (c, f). Green areas = normal lung parenchyma, light blue areas = moderate low-attenuation areas, dark blue areas = severe low-attenuation areas.

Figure 4b.

Figure 4b. Texture-based CT quantification. Coronal thin-section CT images (a, b, d, e) and glyphs (c, f) show the various patterns of emphysema (arrows in a, b, d, e), such as upper lobe–predominant centrilobular emphysema (a–c) and lower lobe–predominant pan-lobular emphysema (d–f), without (a, d) and with (b, e) texture-based quantification overlay and corresponding glyphs (c, f). Green areas = normal lung parenchyma, light blue areas = moderate low-attenuation areas, dark blue areas = severe low-attenuation areas.

Figure 4c.

Figure 4c. Texture-based CT quantification. Coronal thin-section CT images (a, b, d, e) and glyphs (c, f) show the various patterns of emphysema (arrows in a, b, d, e), such as upper lobe–predominant centrilobular emphysema (a–c) and lower lobe–predominant pan-lobular emphysema (d–f), without (a, d) and with (b, e) texture-based quantification overlay and corresponding glyphs (c, f). Green areas = normal lung parenchyma, light blue areas = moderate low-attenuation areas, dark blue areas = severe low-attenuation areas.

Figure 4d.

Figure 4d. Texture-based CT quantification. Coronal thin-section CT images (a, b, d, e) and glyphs (c, f) show the various patterns of emphysema (arrows in a, b, d, e), such as upper lobe–predominant centrilobular emphysema (a–c) and lower lobe–predominant pan-lobular emphysema (d–f), without (a, d) and with (b, e) texture-based quantification overlay and corresponding glyphs (c, f). Green areas = normal lung parenchyma, light blue areas = moderate low-attenuation areas, dark blue areas = severe low-attenuation areas.

Figure 4e.

Figure 4e. Texture-based CT quantification. Coronal thin-section CT images (a, b, d, e) and glyphs (c, f) show the various patterns of emphysema (arrows in a, b, d, e), such as upper lobe–predominant centrilobular emphysema (a–c) and lower lobe–predominant pan-lobular emphysema (d–f), without (a, d) and with (b, e) texture-based quantification overlay and corresponding glyphs (c, f). Green areas = normal lung parenchyma, light blue areas = moderate low-attenuation areas, dark blue areas = severe low-attenuation areas.

Figure 4f.

Figure 4f. Texture-based CT quantification. Coronal thin-section CT images (a, b, d, e) and glyphs (c, f) show the various patterns of emphysema (arrows in a, b, d, e), such as upper lobe–predominant centrilobular emphysema (a–c) and lower lobe–predominant pan-lobular emphysema (d–f), without (a, d) and with (b, e) texture-based quantification overlay and corresponding glyphs (c, f). Green areas = normal lung parenchyma, light blue areas = moderate low-attenuation areas, dark blue areas = severe low-attenuation areas.

Idiopathic Pulmonary Fibrosis

IPF has an unpredictable clinical course, and thin-section CT plays an important role in assessing disease severity and predicting disease progression. Unlike emphysema, in which disease progression can be directly reflected in decreased lung density and can be quantified with attenuation-based threshold measures, quantifying fibrotic changes is relatively more challenging. One whole-lung histogram-based approach focuses on kurtosis, or the sharpness of the peak of a histogram of attenuation distribution of the lungs. Normally, the peak should be sharp, reflecting parenchymal homogeneity. In IPF, the high-attenuation fibrosis and low-attenuation destructive change of honeycombing and reticulations will increase both tails of the histogram and reduce the peak, resulting in abnormally low kurtosis.

Kurtosis has been shown to be predictive of short-term mortality in patients with IPF (38). However, kurtosis was not sensitive to temporal changes in reticular extent and may not serve as a sensitive marker for disease progression (23).

Several texture analysis–based approaches have been developed that directly quantify the extent of fibrotic reticular change (Fig 5). Longitudinal comparison of quantitatively derived fibrosis scores can serve as a useful tool for identifying changes in disease state and for tailoring therapy.

Figure 5a.

Figure 5a. Quantifying longitudinal change. Serial coronal thin-section CT images and corresponding glyphs in a patient with idiopathic pulmonary fibrosis, obtained at presentation (a–c) and at 2-year follow-up (d–f), show an increase in ground-glass opacities (yellow areas), reticular change (orange areas), and honeycombing (red areas), when comparing c and f. The ground-glass opacities represent fine reticulations and fibrosis. Coronal CT images with texture analysis overlay (b, e) show progressive disease. Coronal thin-section CT images without overlay (a, d) show the increase in peripheral reticulations more clearly. Green areas = normal lung parenchyma.

Figure 5b.

Figure 5b. Quantifying longitudinal change. Serial coronal thin-section CT images and corresponding glyphs in a patient with idiopathic pulmonary fibrosis, obtained at presentation (a–c) and at 2-year follow-up (d–f), show an increase in ground-glass opacities (yellow areas), reticular change (orange areas), and honeycombing (red areas), when comparing c and f. The ground-glass opacities represent fine reticulations and fibrosis. Coronal CT images with texture analysis overlay (b, e) show progressive disease. Coronal thin-section CT images without overlay (a, d) show the increase in peripheral reticulations more clearly. Green areas = normal lung parenchyma.

Figure 5c.

Figure 5c. Quantifying longitudinal change. Serial coronal thin-section CT images and corresponding glyphs in a patient with idiopathic pulmonary fibrosis, obtained at presentation (a–c) and at 2-year follow-up (d–f), show an increase in ground-glass opacities (yellow areas), reticular change (orange areas), and honeycombing (red areas), when comparing c and f. The ground-glass opacities represent fine reticulations and fibrosis. Coronal CT images with texture analysis overlay (b, e) show progressive disease. Coronal thin-section CT images without overlay (a, d) show the increase in peripheral reticulations more clearly. Green areas = normal lung parenchyma.

Figure 5d.

Figure 5d. Quantifying longitudinal change. Serial coronal thin-section CT images and corresponding glyphs in a patient with idiopathic pulmonary fibrosis, obtained at presentation (a–c) and at 2-year follow-up (d–f), show an increase in ground-glass opacities (yellow areas), reticular change (orange areas), and honeycombing (red areas), when comparing c and f. The ground-glass opacities represent fine reticulations and fibrosis. Coronal CT images with texture analysis overlay (b, e) show progressive disease. Coronal thin-section CT images without overlay (a, d) show the increase in peripheral reticulations more clearly. Green areas = normal lung parenchyma.

Figure 5e.

Figure 5e. Quantifying longitudinal change. Serial coronal thin-section CT images and corresponding glyphs in a patient with idiopathic pulmonary fibrosis, obtained at presentation (a–c) and at 2-year follow-up (d–f), show an increase in ground-glass opacities (yellow areas), reticular change (orange areas), and honeycombing (red areas), when comparing c and f. The ground-glass opacities represent fine reticulations and fibrosis. Coronal CT images with texture analysis overlay (b, e) show progressive disease. Coronal thin-section CT images without overlay (a, d) show the increase in peripheral reticulations more clearly. Green areas = normal lung parenchyma.

Figure 5f.

Figure 5f. Quantifying longitudinal change. Serial coronal thin-section CT images and corresponding glyphs in a patient with idiopathic pulmonary fibrosis, obtained at presentation (a–c) and at 2-year follow-up (d–f), show an increase in ground-glass opacities (yellow areas), reticular change (orange areas), and honeycombing (red areas), when comparing c and f. The ground-glass opacities represent fine reticulations and fibrosis. Coronal CT images with texture analysis overlay (b, e) show progressive disease. Coronal thin-section CT images without overlay (a, d) show the increase in peripheral reticulations more clearly. Green areas = normal lung parenchyma.

Kim et al (23) developed a supervised classification algorithm using a support vector model (39) to identify different lung parenchymal patterns from which a quantitative lung fibrosis (QLF) score (described previously), which includes only reticular change, and a quantitative ILD (QILD) score, which includes reticular, honeycombing, and ground-glass opacity changes, can be derived. Changes in the QLF score were associated with changes in pulmonary function. The DTA tool, which is based on an unsupervised machine learning model, also demonstrated that quantitative CT-derived fibrosis scores were significantly correlated with pulmonary function at baseline and changes at 15-month follow-up (19). Maldonado et al (40) demonstrated that quantitative short-term longitudinal changes in disease extent as measured by the texture analysis tool CALIPER were predictive of survival.

Another texture-analysis–based automated classification system developed by Park et al (22) used a supervised machine learning approach to demonstrate that a threshold of 22.05% or less of reticular opacity had a high negative predictive value for decline in pulmonary function. This may serve as a useful way to screen for an increased propensity for disease progression, both for determining when a more aggressive management approach may be necessary and for identifying suitable participants in clinical trials.

Chronic Hypersensitivity Pneumonitis

Quantitative techniques have also been shown to be useful in diseases with restrictive, fibrotic, and obstructive components such as hypersensitivity pneumonitis, particularly in the chronic form. Clinically, there is interest in parsing out the subtypes of CHP that have a rapid progression similar to that of IPF. The severity of imaging patterns such as traction bronchiectasis and honeycombing have been shown to better predict mortality than PFTs (41), and the extent of fibrosis on images can identify patients with poor prognoses (42).

Unlike emphysema, which manifests mainly as increased low-attenuation areas, CHP is challenging to quantify with whole-lung summary indices because of its more heterogeneous presentation with honeycombing, reticulation, and mosaic parenchymal patterns. For instance, while changes in lung density correlate to the extent of alveolar destruction in emphysema, disease progression is less easily captured in CHP as the high-attenuated fibrotic and low-attenuated cystic areas may average each other out. Quantitative thin-section CT indices that more accurately reflect disease include the percentage of high-attenuation areas (>–200 HU), the percentage of low-attenuation areas (<–960 HU), and kurtosis (43).

Higher-order statistics texture analysis tools are also well suited to capture the heterogeneous nature of CHP. Jacob et al (44) demonstrated that quantitative CT analysis–based models using CALIPER are comparable to PFT–based models for CHP prognostication and that the CALIPER-calculated extent of fibrosis was an independent predictor of mortality.

Connective Tissue–related ILD

Quantitative thin-section CT–derived parameters may serve as valuable tools for surveillance of disease progression and treatment response because of their reproducibility. Quantifying disease severity with a continuous index rather than a broad visually estimated category can increase our power to detect subtle changes associated with treatment. This has been studied in the setting of scleroderma-related ILD. Kim et al (45) demonstrated that a texture-based measurement reflecting the percentage of lung fibrosis, the QLF score (described previously), had a strong association with changes in forced vital capacity in patients in the Scleroderma Lung Study. The QLF score analysis supported the efficacy of cyclophosphamide over placebo for this indication, which mirrored the clinical trial findings based on forced vital capacity.

In patients with idiopathic inflammatory myopathy–related ILD, CALIPER-derived quantitative measures of disease severity have been shown to correlate with PFT findings at baseline and 1-year follow-up (46). The quantitative features of ILD and the rate of change have also been shown to be useful for differentiating connective tissue–related nonspecific interstitial pneumonitis from UIP and are predictive of overall survival (47).

Lymphangiomyomatosis

LAM is a rare cystic lung disease characterized by thin-walled parenchymal cysts and proliferation of abnormal smooth muscle–like cells, with normal intervening parenchyma. In patients with LAM, the relative area under –950 HU correlated with the results of PFTs (48). Pulmonary cyst volume, as derived from an automated quantitative algorithm that segments out the lung attenuation histogram into two underlying distributions—for cysts and normal parenchyma—also correlated with PFT results (49). Another technique, based on a watershed algorithm, quantifies the amount, distribution, and boundary of cysts. A watershed point is a voxel of relative greatest density around which the density decreases in multiple directions, much like water running off a peak. These quantitative CT features demonstrated improved predictive value of PFTs compared with previous threshold measures (50).

Quantitative analysis has also been used to better elucidate treatment and disease mechanisms. For example, quantitative analysis using the watershed algorithm of thin-section CT data from the Multicenter International Lymphangioleiomyomatosis Efficacy and Safety of Sirolimus (MILES) trial, a double-blind randomized clinical trial that determined that sirolimus stabilizes pulmonary function decline, suggests that this effect may have been mediated by improved expiratory cyst emptying and decreased air trapping (51). In addition, small pulmonary vessel density may serve as a quantitative method to distinguish between LAM and COPD; small pulmonary vessel density was greater in LAM than in COPD, which may reflect differences underlying the pathophysiology in these cystic lung diseases (52).

Other and Unclassifiable Diffuse Lung Diseases

It is not uncommon for the specific pathologic diagnosis of parenchymal disease to be uncertain or for patients to have a combination of obstructive and restrictive diseases. For example, thin-section CT images in patients who smoke can demonstrate some areas of increased lung density owing to inflammation from respiratory bronchiolitis–associated ILD and interstitial fibrosis and other areas of abnormal lucency due to air trapping or emphysema. A similar pattern can also be depicted in hypersensitivity pneumonitis. This mixed physiology can result in confusing physiologic test results, and heterogeneous parenchymal involvement makes the visual assessment of extent and change over time difficult to assess.

Some admixed pulmonary parenchymal patterns remain challenging for textural analysis tools to classify. For example, in patients with concomitant pulmonary fibrosis and emphysema, visual emphysema scores were found to be superior to those of texture analysis–derived scores when assessed against mortality or diffusion capacity of lungs for carbon monoxide. Distinguishing between honeycombing and emphysematous changes in areas with admixed patterns was particularly challenging for the automated tools (53). However, texture analysis holds promise in separately characterizing and quantifying fibrotic components, areas of lucency, and areas of spared lung. Specifically, these quantitative measures in combined pulmonary fibrosis and emphysema can be important for the prediction of survival (53,54).

Even for parenchymal disease that is not classifiable to a specific diagnosis, its quantitative CT features combined with physiologic changes can be powerful predictors of prognosis. Surprisingly, on evaluation of longitudinal variables in a cohort with unclassifiable ILD, increasing CALIPER fibrosis extent was the strongest outcome predictor and remained so even with adjustment for baseline disease severity and when functional vital capacity declines were marginal (55).

Challenges and Limitations

Although quantitative CT of diffuse lung disease holds much promise, there are several pitfalls in interpretation and implementation. Variations in thin-section CT acquisition technique, technical limitations of quantitative analysis tools, and validation of utility in clinical practice remain challenges for quantitative imaging.

Variations in CT Data

Inherent in the quantitative nature of these tools is a reliance on the quality of input data. Variation in the technical parameters of CT, such as reconstruction kernel, section thickness, dose, and the specific manufacturer and model of the CT scanner, will change the signal-to-noise ratio of the resultant images and render interpretation more challenging. Standardized imaging protocols must be used for accurate and comparable quantitative analysis across different centers for clinical trials and across different points of time for each patient (56). For instance, it has been shown that the CT section thickness and reconstruction kernel have a significant effect on the CT quantification of emphysema. This effect varied in magnitude on the basis of the severity of disease (33). Other radiomics features have been shown to differ significantly depending on iterative reconstruction algorithm, pixel size, and CT scanner used (57). Efforts by the Subpopulations and Intermediate Outcome Measures in COPD (SPIROMICS) study, the Radiological Society of North America (RSNA) Quantitative Imaging Biomarker Alliance (QIBA) (58), and the Association of University Radiologists (AUR) Research Alliance Quantitative Imaging Task Force (59) are underway to standardize quantitative thin-section CT protocols (60).

Inspiration

Variation in inspiratory effort can lead to artificial changes in lung thin-section CT attenuation and lung volume. Poor inspiration can result in misinterpretation of areas of atelectasis as increased ground-glass opacities and progression of disease. This can affect both threshold- and histogram-based quantification. Pulmonary attenuation will be artificially inflated for threshold analysis, and the change in parenchymal texture may result in mislabeling by a texture analysis–based algorithm (Fig 6). Inadequate lung inflation may lead to underestimation of emphysema (61). Spirometric-assisted CT acquisition or specific inspiration coaching may minimize some of this variability, but assurance of maximal inspiration or expiration for image acquisition remains a considerable real-world challenge for quantitative pulmonary imaging.

Figure 6a.

Figure 6a. Quantification pitfall: the effect of inspiratory effort. Axial CT images (a, c) with their respective companion texture analysis overlay images (b, d) in the same patient obtained in two examinations performed on the same day with normal (a, b) and decreased (c, d) inspiratory effort show atelectasis from poor inspiration erroneously characterized as ground-glass (yellow areas) and reticular (orange areas) opacities. Green areas = normal lung parenchyma.

Figure 6b.

Figure 6b. Quantification pitfall: the effect of inspiratory effort. Axial CT images (a, c) with their respective companion texture analysis overlay images (b, d) in the same patient obtained in two examinations performed on the same day with normal (a, b) and decreased (c, d) inspiratory effort show atelectasis from poor inspiration erroneously characterized as ground-glass (yellow areas) and reticular (orange areas) opacities. Green areas = normal lung parenchyma.

Figure 6c.

Figure 6c. Quantification pitfall: the effect of inspiratory effort. Axial CT images (a, c) with their respective companion texture analysis overlay images (b, d) in the same patient obtained in two examinations performed on the same day with normal (a, b) and decreased (c, d) inspiratory effort show atelectasis from poor inspiration erroneously characterized as ground-glass (yellow areas) and reticular (orange areas) opacities. Green areas = normal lung parenchyma.

Figure 6d.

Figure 6d. Quantification pitfall: the effect of inspiratory effort. Axial CT images (a, c) with their respective companion texture analysis overlay images (b, d) in the same patient obtained in two examinations performed on the same day with normal (a, b) and decreased (c, d) inspiratory effort show atelectasis from poor inspiration erroneously characterized as ground-glass (yellow areas) and reticular (orange areas) opacities. Green areas = normal lung parenchyma.

Inadequate Extraction

To isolate the lung parenchyma for analysis, all the other thoracic structures need to be removed, including the airways, gastrointestinal tract, vasculature, and rib cage. Multiple segmentation methods are available, ranging from semi- to fully automated techniques (62). As the lungs are predominantly air-filled structures that are significantly different in density compared with that of the surrounding chest wall and mediastinum, extraction by thresholding can be straightforward. However, in the setting of disease in which lung density is altered or with the presence of anatomic variations such as an air-filled patulous esophagus, hiatal hernia, or colonic interposition, simple thresholding and region growing may include nonlung structures or inadequately extracted lungs.

Similarly, region-growing techniques with detection of connected components can be used to optimize extraction of branching structures such as the tracheobronchial tree, but small structures such as the subsegmental bronchioles, regions of severe bronchiectasis, or airways in areas of honeycombing may be exceedingly challenging to correctly extract, particularly in the setting of emphysema, in which the lungs have abnormally low density, or COPD, in which the airways are narrowed or peripherally constricted.

Most airway segmentation methods are dependent on the contrast between the lower-attenuation lung parenchyma and higher-attenuation airway walls. Therefore, extensive parenchymal abnormalities, such as severe emphysema, which can increase the similarity of lung parenchyma to the intraluminal airway lumen, may result in increased leakage or growth of the airway segmentation algorithm outside of its natural walls (62). Such segmentation leaks are especially pronounced for small airways owing to limited resolution and partial volume effects (63). If the airways are not fully removed, they may be misinterpreted as areas of honeycombing or low attenuation (Fig 7). Motion artifacts from cardiac or respiratory motion, metallic beam hardening artifacts, and noisy images from low-radiation-dose scanning and/or a large chest wall can also render segmentation more challenging.

Figure 7a.

Figure 7a. Quantification pitfall: airway underextraction. Accurate lung characterization is reliant on successful extraction of extraparenchymal structures. Gray-scale (a) and texture analysis overlay (b) CT images show that underextracted branching airways (arrow; dark red areas in b) were incorrectly characterized as areas of honeycombing by the automated quantitative tool. The yellow overlay represents areas of ground-glass opacities, the orange overlay represents areas of reticular change, and the green overlay represents normal lung parenchyma.

Figure 7b.

Figure 7b. Quantification pitfall: airway underextraction. Accurate lung characterization is reliant on successful extraction of extraparenchymal structures. Gray-scale (a) and texture analysis overlay (b) CT images show that underextracted branching airways (arrow; dark red areas in b) were incorrectly characterized as areas of honeycombing by the automated quantitative tool. The yellow overlay represents areas of ground-glass opacities, the orange overlay represents areas of reticular change, and the green overlay represents normal lung parenchyma.

Although it would seem relatively straightforward to use the inherent differences between the relatively higher-attenuation vessels and the relatively lower-attenuation lung parenchyma for vessel segmentation, the arterial and venous pulmonary vascular trees are highly complex. A region-growing or multiscale Hessian-based approach to segmentation of lung vessels in a three-dimensional dataset (64,65) can be used to extract these branching structures, but accurate vascular segmentation is exceedingly difficult, even with advanced algorithms and artificial intelligence approaches. Failure to segment small pulmonary vessels might contribute to false interpretation of reticulations. Conversely, mislabeling parenchymal reticular structures as vessels could lead to underestimation of parenchymal disease.

These methods can be further confounded by parenchymal disease. The overlap of branching structures makes differentiation of arterial from venous structures nearly impossible, particularly in non–contrast material–enhanced images. Even with these limitations, however, whole-lung volume and regional characteristics of vascular structures, which are visually difficult to assess, appear to be potentially important quantitative features (44,66,67).

Misinterpretation

Unlike a radiologist who is trained to recognize a broad swath of thoracic pathologic conditions from malignancy to congenital malformations, current algorithms are much narrower in scope. An algorithm can only distinguish among the narrow spectrum of diseases it was specifically trained in and will not be able to correctly diagnose a novel disease. For example, if a texture-based tool trained on ILD was given a case of LAM, a disease it was not trained to recognize, the cysts characteristic of the disease may be misinterpreted as low-attenuation areas suggestive of emphysema (Fig 8). Similarly, if an emphysema index were calculated for such a case, it may erroneously suggest advanced parenchymal destruction. Recognition of the limited scope of quantitative tools and their appropriate application is essential.

Figure 8a.

Figure 8a. Quantification pitfall: limitations of scope. Axial (a) and color-coded overlay (b) CT images show characteristic cysts (arrows) in LAM, a disease that CALIPER was not trained to recognize as low-attenuation areas (dark blue areas) without differentiating them from other low-attenuation areas such as those depicted in emphysema. Parenchymal patterns and diseases must be in the training set, and the algorithms used for characterization need to differentiate these patterns. Green areas = normal lung parenchyma, green outline = left lung, light blue area = mild low-attenuation areas, red outline = right lung, yellow outline = airway.

Figure 8b.

Figure 8b. Quantification pitfall: limitations of scope. Axial (a) and color-coded overlay (b) CT images show characteristic cysts (arrows) in LAM, a disease that CALIPER was not trained to recognize as low-attenuation areas (dark blue areas) without differentiating them from other low-attenuation areas such as those depicted in emphysema. Parenchymal patterns and diseases must be in the training set, and the algorithms used for characterization need to differentiate these patterns. Green areas = normal lung parenchyma, green outline = left lung, light blue area = mild low-attenuation areas, red outline = right lung, yellow outline = airway.

Future Directions

Many current quantitative thin-section CT tools seek to capture known parenchymal patterns with a clear visual imaging correlate or semantic features. As we move further into the realm of machine learning and artificial intelligence, novel texture-based variables without clear visual correlates, or agnostic features, are being discovered that may have practical clinical applications. One example is pulmonary vessel volume (PVV). In several diseases, the vascular structures extracted algorithmically appear to have independent predictive value with regard to mortality and can potentially predict future pulmonary functional decline better than parenchymal features. PVV represents the sum of the volume of pulmonary arteries and veins at the hilum, after the vessels have been excluded, a feature that does not have a visual or linguistic correlate—literally an agnostic feature that was not previously known or labeled (58). The PVV has been shown to be an independent predictor of mortality and pulmonary function decline in CHP and IPF (42,60). The precise relationship of PVV with the underlying disease is unclear; it may reflect underlying disease aggressiveness in some way or vascular changes secondary to disease-related increased negative intrathoracic pressure.

There has been tremendous progress in machine learning techniques and processing power with graphic processing unit–accelerated or dedicated artificial neural network hardware and software applied to advance artificial intelligence approaches to medical imaging. Both for unsupervised data mining and discovery and for more advanced supervised detection and characterization of visual features, the landscape of decision support for radiologists is rapidly changing. It is likely that as quantitative analysis expands to serve more areas, many more agnostic features that humans do not fully perceive or comprehend will be found to be clinically relevant. While multiple quantitative tools are available for diseases such as COPD, UIP, and IPF, studies for other lung diseases such as Langherhan cell histiocytosis or sarcoidosis are limited, and future research in these areas might be of value.

As with any proposed quantitative tool, indices of disease severity must be rigorously validated with clinical measures of disease (6). Similarly, different algorithms should be tested and compared to assess comparative effectiveness. The availability of these programs clinically is also limited currently. Although a few of these programs, such as CALIPER (or Lung Texture Analysis) and Pulmonary Analysis Software Suite and Emphysema Profiler (VIDA Diagnostics, Iowa City, Iowa), are commercially available, many other quantitative tools remain primarily in the research arena.

Conclusion

Quantitative CT analysis is a powerful tool in the evaluation of diffuse lung disease. These models provide reproducible quantitative measures of disease severity and pulmonary parenchymal patterns that facilitate accurate diagnosis and provide an objective standard for evaluating treatment response, disease progression, and disease stratification. With further refinement of quantitative thin-section CT tools, quantitative CT analysis is developing into an increasingly valuable component of diffuse lung disease management.

Disclosures of Conflicts of Interest.— A.C. Activities related to the present article: institution receives royalties for CALIPER software from Imbio. Activities not related to the present article: disclosed no relevant relationships. Other activities: disclosed no relevant relationships. R.A.K. Activities related to the present article: royalties for CALIPER software from Imbio. Activities not related to the present article: disclosed no relevant relationships. Other activities: disclosed no relevant relationships. D.S.G. Activities related to the present article: disclosed no relevant relationships. Activities not related to the present article: received compensation from UCSF to score CT scans for an NIH-funded study; co-inventor on patent application submitted by UCSF titled “Airway muscus impaction.” Other activities: disclosed no relevant relationships. B.J.B. Activities related to the present article: royalties for CALIPER software from Imbio, LLC. Activities not related to the present article: scientific advisor fees from Promedior; institutional grant from NIH/NHLBI for the Radiology Center for the Lung Tissue Research Consortium, for which he is the Principal Investigator; institutional fees from Boeringer Ingelheim for service as a scientific advisor; institutional patent pending for systems and methods for analyzing in vivo tissue volumes using medical imaging data. Other activities: disclosed no relevant relationships. C.W.K. Activities related to the present article: institution receives royalties for CALIPER software from Imbio. Activities not related to the present article: disclosed no relevant relationships. Other activities: disclosed no relevant relationships.

Acknowledgments

The authors would like to thank Lucy Bahn, PhD, and Sonia Watson, PhD, for their help with manuscript preparation.

Presented as an education exhibit at the 2018 RSNA Annual Meeting.

For this journal-based SA-CME, the authors A.C., R.A.K, D.S.G, B.J.B., and C.W.K. have provided disclosures the editor and the reviewers have disclosed no relevant relationships.

References

  • 1. Travis WD, Costabel U, Hansell DM, et al. An official American Thoracic Society/European Respiratory Society statement: Update of the international multidisciplinary classification of the idiopathic interstitial pneumonias. Am J Respir Crit Care Med 2013;188(6):733–748. Crossref, MedlineGoogle Scholar
  • 2. Raghu G, Collard HR, Egan JJ, et al. An official ATS/ERS/JRS/ALAT statement: idiopathic pulmonary fibrosis: evidence-based guidelines for diagnosis and management. Am J Respir Crit Care Med 2011;183(6):788–824. Crossref, MedlineGoogle Scholar
  • 3. Watadani T, Sakai F, Johkoh T, et al. Interobserver variability in the CT assessment of honeycombing in the lungs. Radiology 2013;266(3):936–944. LinkGoogle Scholar
  • 4. Hersh CP, Washko GR, Jacobson FL, et al. Interobserver variability in the determination of upper lobe-predominant emphysema. Chest 2007;131(2):424–431. Crossref, MedlineGoogle Scholar
  • 5. Flaherty KR, Andrei AC, King TE Jr, et al. Idiopathic interstitial pneumonia: do community and academic physicians agree on diagnosis? Am J Respir Crit Care Med 2007;175(10):1054–1060. Crossref, MedlineGoogle Scholar
  • 6. Hansell DM, Goldin JG, King TE Jr, Lynch DA, Richeldi L, Wells AU. CT staging and monitoring of fibrotic interstitial lung diseases in clinical practice and treatment trials: a position paper from the Fleischner Society. Lancet Respir Med 2015;3(6):483–496. Crossref, MedlineGoogle Scholar
  • 7. Raghu G, Collard HR, Anstrom KJ, et al. Idiopathic pulmonary fibrosis: clinically meaningful primary endpoints in phase 3 clinical trials. Am J Respir Crit Care Med 2012;185(10):1044–1048. Crossref, MedlineGoogle Scholar
  • 8. Bankier AA, De Maertelaer V, Keyzer C, Gevenois PA. Pulmonary emphysema: subjective visual grading versus objective quantification with macroscopic morphometry and thin-section CT densitometry. Radiology 1999;211(3):851–858. LinkGoogle Scholar
  • 9. Richeldi L, Collard HR, Jones MG. Idiopathic pulmonary fibrosis. Lancet 2017;389(10082):1941–1952. Crossref, MedlineGoogle Scholar
  • 10. Richeldi L, du Bois RM, Raghu G, et al. Efficacy and safety of nintedanib in idiopathic pulmonary fibrosis. N Engl J Med 2014;370(22):2071–2082. Crossref, MedlineGoogle Scholar
  • 11. King TE Jr, Bradford WZ, Castro-Bernardini S, et al. A phase 3 trial of pirfenidone in patients with idiopathic pulmonary fibrosis. N Engl J Med 2014;370(22):2083–2092. Crossref, MedlineGoogle Scholar
  • 12. Gillies RJ, Kinahan PE, Hricak H. Radiomics: Images Are More than Pictures, They Are Data. Radiology 2016;278(2):563–577. LinkGoogle Scholar
  • 13. Pompe E, van Rikxoort EM, Schmidt M, et al. Parametric response mapping adds value to current computed tomography biomarkers in diagnosing chronic obstructive pulmonary disease. Am J Respir Crit Care Med 2015;191(9):1084–1086. Crossref, MedlineGoogle Scholar
  • 14. Lubner MG, Smith AD, Sandrasegaran K, Sahani DV, Pickhardt PJ. CT Texture Analysis: Definitions, Applications, Biologic Correlates, and Challenges. RadioGraphics 2017;37(5):1483–1503. LinkGoogle Scholar
  • 15. Goldin JG, Kim GHJ, Tseng CH, et al. Longitudinal Changes in Quantitative Interstitial Lung Disease on Computed Tomography after Immunosuppression in the Scleroderma Lung Study II. Ann Am Thorac Soc 2018;15(11):1286–1295. Crossref, MedlineGoogle Scholar
  • 16. VIDA. VIDA Diagnostics Inc. https://vidalung.ai/. Accessed April 1, 2019. Google Scholar
  • 17. Matsuoka S, Yamashiro T, Washko GR, Kurihara Y, Nakajima Y, Hatabu H. Quantitative CT assessment of chronic obstructive pulmonary disease. RadioGraphics 2010;30(1):55–66. LinkGoogle Scholar
  • 18. Sørensen L, Nielsen M, Lo P, Ashraf H, Pedersen JH, de Bruijne M. Texture-based analysis of COPD: a data-driven approach. IEEE Trans Med Imaging 2012;31(1):70–78. Crossref, MedlineGoogle Scholar
  • 19. Humphries SM, Yagihashi K, Huckleberry J, et al. Idiopathic Pulmonary Fibrosis: Data-driven Textural Analysis of Extent of Fibrosis at Baseline and 15-Month Follow-up. Radiology 2017;285(1):270–278. LinkGoogle Scholar
  • 20. Uppaluri R, Hoffman EA, Sonka M, Hunninghake GW, McLennan G. Interstitial lung disease: A quantitative study using the adaptive multiple feature method. Am J Respir Crit Care Med 1999;159(2):519–525. Crossref, MedlineGoogle Scholar
  • 21. Park SC, Tan J, Wang X, et al. Computer-aided detection of early interstitial lung diseases using low-dose CT images. Phys Med Biol 2011;56(4):1139–1153. Crossref, MedlineGoogle Scholar
  • 22. Park HJ, Lee SM, Song JW, et al. Texture-Based Automated Quantitative Assessment of Regional Patterns on Initial CT in Patients With Idiopathic Pulmonary Fibrosis: Relationship to Decline in Forced Vital Capacity. AJR Am J Roentgenol 2016;207(5):976–983. Crossref, MedlineGoogle Scholar
  • 23. Kim HJ, Brown MS, Chong D, et al. Comparison of the quantitative CT imaging biomarkers of idiopathic pulmonary fibrosis at baseline and early change with an interval of 7 months. Acad Radiol 2015;22(1):70–80. Crossref, MedlineGoogle Scholar
  • 24. Erickson BJ, Korfiatis P, Akkus Z, Kline TL. Machine Learning for Medical Imaging. RadioGraphics 2017;37(2):505–515. LinkGoogle Scholar
  • 25. Bartholmai BJ, Raghunath S, Karwoski RA, et al. Quantitative computed tomography imaging of interstitial lung diseases. J Thorac Imaging 2013;28(5):298–307. Crossref, MedlineGoogle Scholar
  • 26. Mohamed Hoesein FA, de Hoop B, Zanen P, et al. CT-quantified emphysema in male heavy smokers: association with lung function decline. Thorax 2011;66(9):782–787. Crossref, MedlineGoogle Scholar
  • 27. Dirksen A, Piitulainen E, Parr DG, et al. Exploring the role of CT densitometry: a randomised study of augmentation therapy in alpha1-antitrypsin deficiency. Eur Respir J 2009;33(6):1345–1353. Crossref, MedlineGoogle Scholar
  • 28. Chapman KR, Burdon JG, Piitulainen E, et al. Intravenous augmentation treatment and lung density in severe a1 antitrypsin deficiency (RAPID): a randomised, double-blind, placebo-controlled trial. Lancet 2015;386(9991):360–368. Crossref, MedlineGoogle Scholar
  • 29. Coxson HO, Dirksen A, Edwards LD, et al. The presence and progression of emphysema in COPD as determined by CT scanning and biomarker expression: a prospective analysis from the ECLIPSE study. Lancet Respir Med 2013;1(2):129–136. Crossref, MedlineGoogle Scholar
  • 30. Heussel CP, Kappes J, Hantusch R, et al. Contrast enhanced CT-scans are not comparable to non-enhanced scans in emphysema quantification. Eur J Radiol 2010;74(3):473–478. Crossref, MedlineGoogle Scholar
  • 31. Schroeder JD, McKenzie AS, Zach JA, et al. Relationships between airflow obstruction and quantitative CT measurements of emphysema, air trapping, and airways in subjects with and without chronic obstructive pulmonary disease. AJR Am J Roentgenol 2013;201(3):W460–W470. Crossref, MedlineGoogle Scholar
  • 32. Busacker A, Newell JD Jr, Keefe T, et al. A multivariate analysis of risk factors for the air-trapping asthmatic phenotype as measured by quantitative CT analysis. Chest 2009;135(1):48–56. Crossref, MedlineGoogle Scholar
  • 33. Gierada DS, Bierhals AJ, Choong CK, et al. Effects of CT section thickness and reconstruction kernel on emphysema quantification relationship to the magnitude of the CT emphysema index. Acad Radiol 2010;17(2):146–156. Crossref, MedlineGoogle Scholar
  • 34. Hochhegger B, Irion KL, Marchiori E. Reconstruction algorithms and CT emphysema measurements. Radiology 2012;263(3):935; author reply 935–936. LinkGoogle Scholar
  • 35. Gierada DS, Guniganti P, Newman BJ, et al. Quantitative CT assessment of emphysema and airways in relation to lung cancer risk. Radiology 2011;261(3):950–959. LinkGoogle Scholar
  • 36. Madani A, Zanen J, de Maertelaer V, Gevenois PA. Pulmonary emphysema: objective quantification at multi-detector row CT—comparison with macroscopic and microscopic morphometry. Radiology 2006;238(3):1036–1043. LinkGoogle Scholar
  • 37. Halper-Stromberg E, Cho MH, Wilson C, et al. Visual Assessment of Chest Computed Tomographic Images Is Independently Useful for Genetic Association Analysis in Studies of Chronic Obstructive Pulmonary Disease. Ann Am Thorac Soc 2017;14(1):33–40. Crossref, MedlineGoogle Scholar
  • 38. Best AC, Meng J, Lynch AM, et al. Idiopathic pulmonary fibrosis: physiologic tests, quantitative CT indexes, and CT visual scores as predictors of mortality. Radiology 2008;246(3):935–940. LinkGoogle Scholar
  • 39. Kim HJ, Li G, Gjertson D, et al. Classification of parenchymal abnormality in scleroderma lung using a novel approach to denoise images collected via a multicenter study. Acad Radiol 2008;15(8):1004–1016. Crossref, MedlineGoogle Scholar
  • 40. Maldonado F, Moua T, Rajagopalan S, et al. Automated quantification of radiological patterns predicts survival in idiopathic pulmonary fibrosis. Eur Respir J 2014; 43(1):204–212. Crossref, MedlineGoogle Scholar
  • 41. Walsh SL, Sverzellati N, Devaraj A, Wells AU, Hansell DM. Chronic hypersensitivity pneumonitis: high resolution computed tomography patterns and pulmonary function indices as prognostic determinants. Eur Radiol 2012;22(8):1672–1679. Crossref, MedlineGoogle Scholar
  • 42. Mooney JJ, Elicker BM, Urbania TH, et al. Radiographic fibrosis score predicts survival in hypersensitivity pneumonitis. Chest 2013;144(2):586–592. Crossref, MedlineGoogle Scholar
  • 43. Tanizawa K, Handa T, Nagai S, et al. Clinical impact of high-attenuation and cystic areas on computed tomography in fibrotic idiopathic interstitial pneumonias. BMC Pulm Med 2015;15(1):74. Crossref, MedlineGoogle Scholar
  • 44. Jacob J, Bartholmai BJ, Rajagopalan S, et al. Mortality prediction in idiopathic pulmonary fibrosis: evaluation of computer-based CT analysis with conventional severity measures. Eur Respir J 2017;49(1):1601011. Crossref, MedlineGoogle Scholar
  • 45. Kim HJ, Brown MS, Elashoff R, et al. Quantitative texture-based assessment of one-year changes in fibrotic reticular patterns on HRCT in scleroderma lung disease treated with oral cyclophosphamide. Eur Radiol 2011;21(12):2455–2465. Crossref, MedlineGoogle Scholar
  • 46. Ungprasert P, Wilton KM, Ernste FC, et al. Novel Assessment of Interstitial Lung Disease Using the “Computer-Aided Lung Informatics for Pathology Evaluation and Rating” (CALIPER) Software System in Idiopathic Inflammatory Myopathies. Hai 2017;195(5):545–552. Google Scholar
  • 47. De Giacomi F, Raghunath S, Karwoski R, Bartholmai BJ, Moua T. Short-term Automated Quantification of Radiologic Changes in the Characterization of Idiopathic Pulmonary Fibrosis Versus Nonspecific Interstitial Pneumonia and Prediction of Long-term Survival. J Thorac Imaging 2018;33(2):124–131. Crossref, MedlineGoogle Scholar
  • 48. Avila NA, Kelly JA, Dwyer AJ, Johnson DL, Jones EC, Moss J. Lymphangioleiomyomatosis: correlation of qualitative and quantitative thin-section CT with pulmonary function tests and assessment of dependence on pleurodesis. Radiology 2002;223(1):189–197. LinkGoogle Scholar
  • 49. Schmithorst VJ, Altes TA, Young LR, et al. Automated algorithm for quantifying the extent of cystic change on volumetric chest CT: initial results in Lymphangioleiomyomatosis. AJR Am J Roentgenol 2009;192(4):1037–1044. Crossref, MedlineGoogle Scholar
  • 50. Lo P, Brown MS, Kim H, et al. Cyst-based measurements for assessing lymphangioleiomyomatosis in computed tomography. Med Phys 2015;42(5):2287–2295. Crossref, MedlineGoogle Scholar
  • 51. Argula RG, Kokosi M, Lo P, et al. A Novel Quantitative Computed Tomographic Analysis Suggests How Sirolimus Stabilizes Progressive Air Trapping in Lymphangioleiomyomatosis. Ann Am Thorac Soc 2016;13(3):342–349. Crossref, MedlineGoogle Scholar
  • 52. Ando K, Tobino K, Kurihara M, et al. Quantitative CT analysis of small pulmonary vessels in lymphangioleiomyomatosis. Eur J Radiol 2012;81(12):3925–3930. Crossref, MedlineGoogle Scholar
  • 53. Jacob J, Bartholmai BJ, Rajagopalan S, et al. Functional and prognostic effects when emphysema complicates idiopathic pulmonary fibrosis. Eur Respir J 2017;50(1):1700379. Crossref, MedlineGoogle Scholar
  • 54. Alsumrain M, De Giacomi F, Nasim F, et al. Combined pulmonary fibrosis and emphysema as a clinicoradiologic entity: Characterization of presenting lung fibrosis and implications for survival. Respir Med 2019;146:106–112. Crossref, MedlineGoogle Scholar
  • 55. Jacob J, Bartholmai BJ, Rajagopalan S, et al. Unclassifiable-interstitial lung disease: Outcome prediction using CT and functional indices. Respir Med 2017;130:43–51. Crossref, MedlineGoogle Scholar
  • 56. Goldin JG. Computed tomography as a biomarker in clinical trials imaging. J Thorac Imaging 2013;28(5):291–297. Crossref, MedlineGoogle Scholar
  • 57. Kim H, Goo JM, Ohno Y, et al. Effect of Reconstruction Parameters on the Quantitative Analysis of Chest Computed Tomography. J Thorac Imaging 2019;34(2):92–102. Crossref, MedlineGoogle Scholar
  • 58. RSNA QIBA. Radiological Society of North America. http://qibawiki.rsna.org/index.php/Lung_Density_Biomarker_Ctte. Updated September 25, 2019. Accessed June 1, 2019. Google Scholar
  • 59. Abramson RG, Burton KR, Yu JP, et al. Methods and challenges in quantitative imaging biomarker development. Acad Radiol 2015;22(1):25–32. Crossref, MedlineGoogle Scholar
  • 60. Sieren JP, Newell JD Jr, Barr RG, et al. SPIROMICS Protocol for Multicenter Quantitative Computed Tomography to Phenotype the Lungs. Am J Respir Crit Care Med 2016;194(7):794–806. Crossref, MedlineGoogle Scholar
  • 61. Madani A, Van Muylem A, Gevenois PA. Pulmonary emphysema: effect of lung volume on objective quantification at thin-section CT. Radiology 2010;257(1):260–268. LinkGoogle Scholar
  • 62. Tschirren J, Hoffman EA, McLennan G, Sonka M. Segmentation and quantitative analysis of intrathoracic airway trees from computed tomography images. Proc Am Thorac Soc 2005;2(6):484–487, 503–504. Crossref, MedlineGoogle Scholar
  • 63. Charbonnier JP, Rikxoort EMV, Setio AAA, Schaefer-Prokop CM, Ginneken BV, Ciompi F. Improving airway segmentation in computed tomography using leak detection with convolutional networks. Med Image Anal 2017;36:52–60. Crossref, MedlineGoogle Scholar
  • 64. Orkisz M, Hoyos MH, Romanello VP, Romanello CP, Prieto JC, Revol-Muller C. Segmentation of the pulmonary vascular trees in 3D CT images using variational region-growing. IRBM 2014;35(1):11–19. CrossrefGoogle Scholar
  • 65. Öksüz I. Multi-Scale Hessian Based Approach for Segmentation of Lung Vessel Tree in 3-D CTA Data. 2012. Google Scholar
  • 66. Jacob J, Bartholmai BJ, Egashira R, et al. Chronic hypersensitivity pneumonitis: identification of key prognostic determinants using automated CT analysis. BMC Pulm Med 2017;17(1):81. Crossref, MedlineGoogle Scholar
  • 67. Jacob J, Bartholmai BJ, Rajagopalan S, et al. Predicting Outcomes in Idiopathic Pulmonary Fibrosis Using Automated Computed Tomographic Analysis. Am J Respir Crit Care Med 2018;198(6):767–776. Crossref, MedlineGoogle Scholar

Article History

Received: Apr 4 2019
Revision requested: May 31 2019
Revision received: June 27 2019
Accepted: July 11 2019
Published online: Nov 29 2019
Published in print: Jan 2020