Disease Staging and Prognosis in Smokers Using Deep Learning in Chest Computed Tomography.

RATIONALE: Deep learning is a powerful tool that may allow for improved outcome prediction. OBJECTIVES: To determine if deep learning, specifically convolutional neural network (CNN) analysis, could detect and stage chronic obstructive pulmonary disease (COPD) and predict acute respiratory disease (ARD) events and mortality in smokers. METHODS: A CNN was trained using computed tomography scans from 7,983 COPDGene participants and evaluated using 1,000 nonoverlapping COPDGene participants and 1,672 ECLIPSE participants. Logistic regression (C statistic and the Hosmer-Lemeshow test) was used to assess COPD diagnosis and ARD prediction. Cox regression (C index and the Greenwood-Nam-D'Agnostino test) was used to assess mortality. MEASUREMENTS AND MAIN RESULTS: In COPDGene, the C statistic for the detection of COPD was 0.856. A total of 51.1% of participants in COPDGene were accurately staged and 74.95% were within one stage. In ECLIPSE, 29.4% were accurately staged and 74.6% were within one stage. In COPDGene and ECLIPSE, the C statistics for ARD events were 0.64 and 0.55, respectively, and the Hosmer-Lemeshow P values were 0.502 and 0.380, respectively, suggesting no evidence of poor calibration. In COPDGene and ECLIPSE, CNN predicted mortality with fair discrimination (C indices, 0.72 and 0.60, respectively), and without evidence of poor calibration (Greenwood-Nam-D'Agnostino P values, 0.307 and 0.331, respectively). CONCLUSIONS: A deep-learning approach that uses only computed tomography imaging data can identify those smokers who have COPD and predict who are most likely to have ARD events and those with the highest mortality. At a population level CNN analysis may be a powerful tool for risk assessment.

Respiratory Symptoms in Young Adults and Future Lung Disease. The CARDIA Lung Study.

RATIONALE: There are limited data on factors in young adulthood that predict future lung disease. OBJECTIVES: To determine the relationship between respiratory symptoms, loss of lung health, and incident respiratory disease in a population-based study of young adults. METHODS: We examined prospective data from 2,749 participants in the CARDIA (Coronary Artery Risk Development in Young Adults) study who completed respiratory symptom questionnaires at baseline and 2 years later and repeated spirometry measurements over 30 years. MEASUREMENTS AND MAIN RESULTS: Cough or phlegm, episodes of bronchitis, wheeze, shortness of breath, and chest illnesses at both baseline and Year 2 were the main predictor variables in models assessing decline in FEV1 and FVC from Year 5 to Year 30, incident obstructive and restrictive lung physiology, and visual emphysema on thoracic computed tomography scan. After adjustment for covariates, including body mass index, asthma, and smoking, report of any symptom was associated with -2.71 ml/yr excess decline in FEV1 (P < 0.001) and -2.18 in FVC (P < 0.001) as well as greater odds of incident (prebronchodilator) obstructive (odds ratio [OR], 1.63; 95% confidence interval [CI], 1.24-2.14) and restrictive (OR, 1.40; 95% CI, 1.09-1.80) physiology. Cough-related symptoms (OR, 1.56; 95% CI, 1.13-2.16) were associated with greater odds of future emphysema. CONCLUSIONS: Persistent respiratory symptoms in young adults are associated with accelerated decline in lung function, incident obstructive and restrictive physiology, and greater odds of future radiographic emphysema.

Interstitial Features at Chest CT Enhance the Deleterious Effects of Emphysema in the COPDGene Cohort.

Purpose To determine if interstitial features at chest CT enhance the effect of emphysema on clinical disease severity in smokers without clinical pulmonary fibrosis. Materials and Methods In this retrospective cohort study, an objective CT analysis tool was used to measure interstitial features (reticular changes, honeycombing, centrilobular nodules, linear scar, nodular changes, subpleural lines, and ground-glass opacities) and emphysema in 8266 participants in a study of chronic obstructive pulmonary disease (COPD) called COPDGene (recruited between October 2006 and January 2011). Additive differences in patients with emphysema with interstitial features and in those without interstitial features were analyzed by using t tests, multivariable linear regression, and Kaplan-Meier analysis. Multivariable linear and Cox regression were used to determine if interstitial features modified the effect of continuously measured emphysema on clinical measures of disease severity and mortality. Results Compared with individuals with emphysema alone, those with emphysema and interstitial features had a higher percentage predicted forced expiratory volume in 1 second (absolute difference, 6.4%; P < .001), a lower percentage predicted diffusing capacity of lung for carbon monoxide (DLCO) (absolute difference, 7.4%; P = .034), a 0.019 higher right ventricular-to-left ventricular (RVLV) volume ratio (P = .029), a 43.2-m shorter 6-minute walk distance (6MWD) (P < .001), a 5.9-point higher St George's Respiratory Questionnaire (SGRQ) score (P < .001), and 82% higher mortality (P < .001). In addition, interstitial features modified the effect of emphysema on percentage predicted DLCO, RVLV volume ratio, 6WMD, SGRQ score, and mortality (P for interaction < .05 for all). Conclusion In smokers, the combined presence of interstitial features and emphysema was associated with worse clinical disease severity and higher mortality than was emphysema alone. In addition, interstitial features enhanced the deleterious effects of emphysema on clinical disease severity and mortality.

Densitometric and local histogram based analysis of computed tomography images in patients with idiopathic pulmonary fibrosis.

BACKGROUND: Prior studies of clinical prognostication in idiopathic pulmonary fibrosis (IPF) using computed tomography (CT) have often used subjective analyses or have evaluated quantitative measures in isolation. This study examined associations between both densitometric and local histogram based quantitative CT measurements with pulmonary function test (PFT) parameters and mortality. In addition, this study sought to compare risk prediction scores that incorporate quantitative CT measures with previously described systems. METHODS: Forty six patients with biopsy proven IPF were identified from a registry of patients with interstitial lung disease at Brigham and Women's Hospital in Boston, MA. CT scans for each subject were visually scored using a previously published method. After a semi-automated method was used to segment the lungs from the surrounding tissue, densitometric measurements including the percent high attenuating area, mean lung density, skewness and kurtosis were made for the entirety of each patient's lungs. A separate, automated tool was used to detect and quantify the percent of lung occupied by interstitial lung features. These analyses were used to create clinical and quantitative CT based risk prediction scores, and the performance of these was compared to the performance of clinical and visual analysis based methods. RESULTS: All of the densitometric measures were correlated with forced vital capacity and diffusing capacity, as were the total amount of interstitial change and the percentage of interstitial change that was honeycombing measured using the local histogram method. Higher percent high attenuating area, higher mean lung density, lower skewness, lower kurtosis and a higher percentage of honeycombing were associated with worse transplant free survival. The quantitative CT based risk prediction scores performed similarly to the clinical and visual analysis based methods. CONCLUSIONS: Both densitometric and feature based quantitative CT measures correlate with pulmonary function test measures and are associated with transplant free survival. These objective measures may be useful for identifying high risk patients and monitoring disease progression. Further work will be needed to validate these measures and the quantitative imaging based risk prediction scores in other cohorts.

Robust breast cancer detection in mammography and digital breast tomosynthesis using an annotation-efficient deep learning approach.

Breast cancer remains a global challenge, causing over 600,000 deaths in 2018 (ref. 1). To achieve earlier cancer detection, health organizations worldwide recommend screening mammography, which is estimated to decrease breast cancer mortality by 20-40% (refs. 2,3). Despite the clear value of screening mammography, significant false positive and false negative rates along with non-uniformities in expert reader availability leave opportunities for improving quality and access4,5. To address these limitations, there has been much recent interest in applying deep learning to mammography6-18, and these efforts have highlighted two key difficulties: obtaining large amounts of annotated training data and ensuring generalization across populations, acquisition equipment and modalities. Here we present an annotation-efficient deep learning approach that (1) achieves state-of-the-art performance in mammogram classification, (2) successfully extends to digital breast tomosynthesis (DBT; '3D mammography'), (3) detects cancers in clinically negative prior mammograms of patients with cancer, (4) generalizes well to a population with low screening rates and (5) outperforms five out of five full-time breast-imaging specialists with an average increase in sensitivity of 14%. By creating new 'maximum suspicion projection' (MSP) images from DBT data, our progressively trained, multiple-instance learning approach effectively trains on DBT exams using only breast-level labels while maintaining localization-based interpretability. Altogether, our results demonstrate promise towards software that can improve the accuracy of and access to screening mammography worldwide.

A simple assessment of lung nodule location for reduction in unnecessary invasive procedures.

BACKGROUND: CT screening for lung cancer results in a significant mortality reduction but is complicated by invasive procedures performed for evaluation of the many detected benign nodules. The purpose of this study was to evaluate measures of nodule location within the lung as predictors of malignancy. METHODS: We analyzed images and data from 3,483 participants in the National Lung Screening Trial (NLST). All nodules (4-20 mm) were characterized by 3D geospatial location using a Cartesian coordinate system and evaluated in logistic regression analysis. Model development and probability cutpoint selection was performed in the NLST testing set. The Geospatial test was then validated in the NLST testing set, and subsequently replicated in a new cohort of 147 participants from The Detection of Early Lung Cancer Among Military Personnel (DECAMP) Consortium. RESULTS: The Geospatial Test, consisting of the superior-inferior distance (Z distance), nodule diameter, and radial distance (carina to nodule) performed well in both the NLST validation set (AUC 0.85) and the DECAMP replication cohort (AUC 0.75). A negative Geospatial Test resulted in a less than 2% risk of cancer across all nodule diameters. The Geospatial Test correctly reclassified 19.7% of indeterminate nodules with a diameter over 6mm as benign, while only incorrectly classifying 1% of cancerous nodules as benign. In contrast, the parsimonious Brock Model applied to the same group of nodules correctly reclassified 64.5% of indeterminate nodules as benign but resulted in misclassification of a cancer as benign in 18.2% of the cases. Applying the Geospatial test would result in reducing invasive procedures performed for benign lesions by 11.3% with a low rate of misclassification (1.3%). In contrast, the Brock model applied to the same group of patients results in decreasing invasive procedures for benign lesion by 39.0% but misclassifying 21.1% of cancers as benign. CONCLUSIONS: Utilizing information about geospatial location within the lung improves risk assessment for indeterminate lung nodules and may reduce unnecessary procedures. TRIAL REGISTRATION: NCT00047385, NCT01785342.

An open-source framework for pulmonary fissure completeness assessment.

We present an open-source framework for pulmonary fissure completeness assessment. Fissure incompleteness has been shown to associate with emphysema treatment outcomes, motivating the development of tools that facilitate completeness estimation. Generally, the task of fissure completeness assessment requires accurate detection of fissures and definition of the boundary surfaces separating the lung lobes. The framework we describe acknowledges a) the modular nature of fissure detection and lung lobe segmentation (lobe boundary detection), and b) that methods to address these challenges are varied and continually developing. It is designed to be readily deployable on existing lung lobe segmentation and fissure detection data sets. The framework consists of multiple components: a flexible quality control module that enables rapid assessment of lung lobe segmentations, an interactive lobe segmentation tool exposed through 3D Slicer for handling challenging cases, a flexible fissure representation using particles-based sampling that can handle fissure feature-strength or binary fissure detection images, and a module that performs fissure completeness estimation using voxel counting and a novel surface area estimation approach. We demonstrate the usage of the proposed framework by deploying on 100 cases exhibiting various levels of fissure completeness. We compare the two completeness level approaches and also compare to visual reads. The code is available to the community via github as part of the Chest Imaging Platform and a 3D Slicer extension module.

Diffeomorphic Lung Registration Using Deep CNNs and Reinforced Learning.

Image registration is a well-known problem in the field of medical imaging. In this paper, we focus on the registration of chest inspiratory and expiratory computed tomography (CT) scans from the same patient. Our method recovers the diffeomorphic elastic displacement vector field (DVF) by jointly regressing the direct and the inverse transformation. Our architecture is based on the RegNet network but we implement a reinforced learning strategy that can accommodate a large training dataset. Our results show that our method performs with a lower estimation error for the same number of epochs than the RegNet approach.

3D Pulmonary Artery Segmentation from CTA Scans Using Deep Learning with Realistic Data Augmentation.

The characterization of the vasculature in the mediastinum, more specifically the pulmonary artery, is of vital importance for the evaluation of several pulmonary vascular diseases. Thus, the goal of this study is to automatically segment the pulmonary artery (PA) from computed tomography angiography images, which opens up the opportunity for more complex analysis of the evolution of the PA geometry in health and disease and can be used in complex fluid mechanics models or individualized medicine. For that purpose, a new 3D convolutional neural network architecture is proposed, which is trained on images coming from different patient cohorts. The network makes use a strong data augmentation paradigm based on realistic deformations generated by applying principal component analysis to the deformation fields obtained from the affine registration of several datasets. The network is validated on 91 datasets by comparing the automatic segmentations with semi-automatically delineated ground truths in terms of mean Dice and Jaccard coefficients and mean distance between surfaces, which yields values of 0.89, 0.80 and 1.25 mm, respectively. Finally, a comparison against a Unet architecture is also included.

Multiorgan structures detection using deep convolutional neural networks.

Many automatic image analysis algorithms in medical imaging require a good initialization to work properly. A similar problem occurs in many imaging-based clinical workflows, which depend on anatomical landmarks. The localization of anatomic structures based on a defined context provides with a solution to that problem, which turns out to be more challenging in medical imaging where labeled images are difficult to obtain. We propose a two-stage process to detect and regress 2D bounding boxes of predefined anatomical structures based on a 2D surrounding context. First, we use a deep convolutional neural network (DCNN) architecture to detect the optimal slice where an anatomical structure is present, based on relevant landmark features. After this detection, we employ a similar architecture to perform a 2D regression with the aim of proposing a bounding box where the structure is encompassed. We trained and tested our system for 57 anatomical structures defined in axial, sagittal and coronal planes with a dataset of 504 labeled Computed Tomography (CT) scans. We compared our method with a well-known object detection algorithm (Viola Jones) and with the inter-rater error for two human experts. Despite the relatively small number of scans and the exhaustive number of structures analyzed, our method obtained promising and consistent results, which proves our architecture very generalizable to other anatomical structures.

Clinical and Genetic Associations of Objectively Identified Interstitial Changes in Smokers.

BACKGROUND: Smoking-related lung injury may manifest on CT scans as both emphysema and interstitial changes. We have developed an automated method to quantify interstitial changes and hypothesized that this measurement would be associated with lung function, quality of life, mortality, and a mucin 5B (MUC5B) polymorphism. METHODS: Using CT scans from the Genetic Epidemiology of COPD Study, we objectively labeled lung parenchyma as a tissue subtype. We calculated the percentage of the lung occupied by interstitial subtypes. RESULTS: A total of 8,345 participants had clinical and CT scanning data available. A 5% absolute increase in interstitial changes was associated with an absolute decrease in FVC % predicted of 2.47% (P < .001) and a 1.36-point higher St. George's Respiratory Questionnaire score (P < .001). Among the 6,827 participants with mortality data, a 5% increase in interstitial changes was associated with a 29% increased risk of death (P < .001). These associations were present in a subgroup without visually defined interstitial lung abnormalities, as well as in those with normal spirometric test results, and in those without chronic respiratory symptoms. In non-Hispanic whites, for each copy of the minor allele of the MUC5B promoter polymorphism, there was a 0.64% (P < .001) absolute increase in the percentage of lung with interstitial changes. CONCLUSIONS: Objective interstitial changes on CT scans were associated with impaired lung function, worse quality of life, increased mortality, and more copies of a MUC5B promoter polymorphism, suggesting that these changes may be a marker of susceptibility to smoking-related lung injury, detectable even in those who are healthy by other measures.

The Objective Identification and Quantification of Interstitial Lung Abnormalities in Smokers.

RATIONALE AND OBJECTIVES: Previous investigation suggests that visually detected interstitial changes in the lung parenchyma of smokers are highly clinically relevant and predict outcomes, including death. Visual subjective analysis to detect these changes is time-consuming, insensitive to subtle changes, and requires training to enhance reproducibility. Objective detection of such changes could provide a method of disease identification without these limitations. The goal of this study was to develop and test a fully automated image processing tool to objectively identify radiographic features associated with interstitial abnormalities in the computed tomography scans of a large cohort of smokers. MATERIALS AND METHODS: An automated tool that uses local histogram analysis combined with distance from the pleural surface was used to detect radiographic features consistent with interstitial lung abnormalities in computed tomography scans from 2257 individuals from the Genetic Epidemiology of COPD study, a longitudinal observational study of smokers. The sensitivity and specificity of this tool was determined based on its ability to detect the visually identified presence of these abnormalities. RESULTS: The tool had a sensitivity of 87.8% and a specificity of 57.5% for the detection of interstitial lung abnormalities, with a c-statistic of 0.82, and was 100% sensitive and 56.7% specific for the detection of the visual subtype of interstitial abnormalities called fibrotic parenchymal abnormalities, with a c-statistic of 0.89. CONCLUSIONS: In smokers, a fully automated image processing tool is able to identify those individuals who have interstitial lung abnormalities with moderate sensitivity and specificity.

Three-dimensional Printing and 3D Slicer: Powerful Tools in Understanding and Treating Structural Lung Disease.

Recent advances in the three-dimensional (3D) printing industry have enabled clinicians to explore the use of 3D printing in preprocedural planning, biomedical tissue modeling, and direct implantable device manufacturing. Despite the increased adoption of rapid prototyping and additive manufacturing techniques in the health-care field, many physicians lack the technical skill set to use this exciting and useful technology. Additionally, the growth in the 3D printing sector brings an ever-increasing number of 3D printers and printable materials. Therefore, it is important for clinicians to keep abreast of this rapidly developing field in order to benefit. In this Ahead of the Curve, we review the history of 3D printing from its inception to the most recent biomedical applications. Additionally, we will address some of the major barriers to wider adoption of the technology in the medical field. Finally, we will provide an initial guide to 3D modeling and printing by demonstrating how to design a personalized airway prosthesis via 3D Slicer. We hope this information will reduce the barriers to use and increase clinician participation in the 3D printing health-care sector.