Emphysema quantification using low-dose computed tomography with deep learning-based kernel conversion comparison

Deep Learning-Based Kernel Adaptation Enhances Quantification of Emphysema on Low-Dose Chest CT for Predicting Long-Term Mortality

OBJECTIVES

The aim of this study was to ascertain the predictive value of quantifying emphysema using low-dose computed tomography (LDCT) post deep learning-based kernel adaptation on long-term mortality.

MATERIALS AND METHODS

This retrospective study investigated LDCTs obtained from asymptomatic individuals aged 60 years or older during health checkups between February 2009 and December 2016. These LDCTs were reconstructed using a 1- or 1.25-mm slice thickness alongside high-frequency kernels. A deep learning algorithm, capable of generating CT images that resemble standard-dose and low-frequency kernel images, was applied to these LDCTs. To quantify emphysema, the lung volume percentage with an attenuation value less than or equal to -950 Hounsfield units (LAA-950) was gauged before and after kernel adaptation. Low-dose chest CTs with LAA-950 exceeding 6% were deemed emphysema-positive according to the Fleischner Society statement. Survival data were sourced from the National Registry Database at the close of 2021. The risk of nonaccidental death, excluding causes such as injury or poisoning, was explored according to the emphysema quantification results using multivariate Cox proportional hazards models.

RESULTS

The study comprised 5178 participants (mean age ± SD, 66 ± 3 years; 3110 males). The median LAA-950 (18.2% vs 2.6%) and the proportion of LDCTs with LAA-950 exceeding 6% (96.3% vs 39.3%) saw a significant decline after kernel adaptation. There was no association between emphysema quantification before kernel adaptation and the risk of nonaccidental death. Nevertheless, after kernel adaptation, higher LAA-950 (hazards ratio for 1% increase, 1.01; P = 0.045) and LAA-950 exceeding 6% (hazards ratio, 1.36; P = 0.008) emerged as independent predictors of nonaccidental death, upon adjusting for age, sex, and smoking status.

CONCLUSIONS

The application of deep learning for kernel adaptation proves instrumental in quantifying pulmonary emphysema on LDCTs, establishing itself as a potential predictive tool for long-term nonaccidental mortality in asymptomatic individuals.

Automated Coronary Artery Calcium and Quantitative Emphysema in Lung Cancer Screening: Association With Mortality, Lung Cancer Incidence, and Airflow Obstruction

PURPOSE

To assess automated coronary artery calcium (CAC) and quantitative emphysema (percentage of low attenuation areas [%LAA]) for predicting mortality and lung cancer (LC) incidence in LC screening. To explore correlations between %LAA, CAC, and forced expiratory value in 1 second (FEV 1 ) and the discriminative ability of %LAA for airflow obstruction.

MATERIALS AND METHODS

Baseline low-dose computed tomography scans of the BioMILD trial were analyzed using an artificial intelligence software. Univariate and multivariate analyses were performed to estimate the predictive value of %LAA and CAC. Harrell C -statistic and time-dependent area under the curve (AUC) were reported for 3 nested models (Model survey : age, sex, pack-years; Model survey-LDCT : Model survey plus %LAA plus CAC; Model final : Model survey-LDCT plus selected confounders). The correlations between %LAA, CAC, and FEV 1 and the discriminative ability of %LAA for airflow obstruction were tested using the Pearson correlation coefficient and AUC-receiver operating characteristic curve, respectively.

RESULTS

A total of 4098 volunteers were enrolled. %LAA and CAC independently predicted 6-year all-cause (Model final hazard ratio [HR], 1.14 per %LAA interquartile range [IQR] increase [95% CI, 1.05-1.23], 2.13 for CAC ≥400 [95% CI, 1.36-3.28]), noncancer (Model final HR, 1.25 per %LAA IQR increase [95% CI, 1.11-1.37], 3.22 for CAC ≥400 [95%CI, 1.62-6.39]), and cardiovascular (Model final HR, 1.25 per %LAA IQR increase [95% CI, 1.00-1.46], 4.66 for CAC ≥400, [95% CI, 1.80-12.58]) mortality, with an increase in concordance probability in Model survey-LDCT compared with Model survey ( P <0.05). No significant association with LC incidence was found after adjustments. Both biomarkers negatively correlated with FEV 1 ( P <0.01). %LAA identified airflow obstruction with a moderate discriminative ability (AUC, 0.738).

CONCLUSIONS

Automated CAC and %LAA added prognostic information to age, sex, and pack-years for predicting mortality but not LC incidence in an LC screening setting. Both biomarkers negatively correlated with FEV 1 , with %LAA enabling the identification of airflow obstruction with moderate discriminative ability.

Deep learning-based harmonization of CT reconstruction kernels towards improved clinical task performance

OBJECTIVES

To develop a deep learning-based harmonization framework, assessing whether it can improve performance of radiomics models given different kernels in different clinical tasks and additionally generalize to mitigate the effects of new/unobserved kernels on radiomics features.

METHODS

Patient data with 2 reconstruction kernels and phantom data with 22 reconstruction kernels were included. Eighty-five patients were studied for lymph node metastasis (LNM) prediction, and 164 patients for differential diagnosis between lung cancer (LC) and pulmonary tuberculosis (TB). Two convolutional neural network (CNN) models were developed to convert images (i) from B70f to B30f (CNNa) and (ii) from B30f to B70f (CNNb). Model performance between the two kernels was evaluated using AUC and compared with other well-known harmonization methods. Patient-normalized feature difference (PNFD) was used to identify the incompatible kernels (i.e., kernel with median PNFD > 1) with baseline (B30f/B70f), and measure the ability of the CNN models to convert the non-comparable kernels.

RESULTS

For LC versus pulmonary TB diagnosis, AUCs of CNNa vs. others were 0.85 vs. 0.54-0.74 (p = 0.0001-0.0003), and for CNNb vs. others: 0.87 vs. 0.54-0.86 (p = 0.0001-0.55). For LNM prediction, AUCs of CNNa vs. others were 0.68 vs. 0.56-0.61 (p = 0.10-0.39), and for CNNb vs. others: 0.78 vs. 0.70-0.73 (p = 0.07-0.40). After CNN harmonization, 17 of 20 (85%) of investigated unknown kernels produced comparable radiomics feature values relative to baseline (median PNFD from 1.10-2.31 to 0.23-1.13).

CONCLUSION

The CNN harmonization effectively improved performance of radiomics models between reconstruction kernels in different clinical tasks, and reduced feature differences between unknown kernels vs. baseline.

KEY POINTS

• The soft (B30f) and sharp (B70f) kernels strongly affect radiomics reproducibility and generalizability. • The convolutional neural network (CNN) harmonization methods performed better than location-scale (ComBat and centering-scaling) and matrix factorization harmonization methods (based on singular value decomposition (SVD) and independent component analysis (ICA)) in both clinical tasks. • The CNN harmonization methods improve feature reproducibility not only between specific kernels (B30f and B70f) from the same scanner, but also between unobserved kernels from different scanners of different vendors.

Fully automatic coronary calcium scoring in non-ECG-gated low-dose chest CT: comparison with ECG-gated cardiac CT

OBJECTIVES

To validate an artificial intelligence (AI)-based fully automatic coronary artery calcium (CAC) scoring system on non-electrocardiogram (ECG)-gated low-dose chest computed tomography (LDCT) using multi-institutional datasets with manual CAC scoring as the reference standard.

METHODS

This retrospective study included 452 subjects from three academic institutions, who underwent both ECG-gated calcium scoring computed tomography (CSCT) and LDCT scans. For all CSCT and LDCT scans, automatic CAC scoring (CAC_auto) was performed using AI-based software, and manual CAC scoring (CAC_man) was set as the reference standard. The reliability and agreement of CAC_auto was evaluated and compared with that of CAC_man using intraclass correlation coefficients (ICCs) and Bland-Altman plots. The reliability between CAC_auto and CAC_man for CAC severity categories was analyzed using weighted kappa (κ) statistics.

RESULTS

CAC_auto on CSCT and LDCT yielded a high ICC (0.998, 95% confidence interval (CI) 0.998-0.999 and 0.989, 95% CI 0.987-0.991, respectively) and a mean difference with 95% limits of agreement of 1.3 ± 37.1 and 0.8 ± 75.7, respectively. CAC_auto achieved excellent reliability for CAC severity (κ = 0.918-0.972) on CSCT and good to excellent but heterogenous reliability among datasets (κ = 0.748-0.924) on LDCT.

CONCLUSIONS

The application of an AI-based automatic CAC scoring software to LDCT shows good to excellent reliability in CAC score and CAC severity categorization in multi-institutional datasets; however, the reliability varies among institutions.

KEY POINTS

• AI-based automatic CAC scoring on LDCT shows excellent reliability with manual CAC scoring in multi-institutional datasets. • The reliability for CAC score-based severity categorization varies among datasets. • Automatic scoring for LDCT shows a higher false-positive rate than automatic scoring for CSCT, and most common causes of a false-positive are image noise and artifacts for both CSCT and LDCT.

A Challenge for Emphysema Quantification Using a Deep Learning Algorithm With Low-dose Chest Computed Tomography

PURPOSE

We aimed to identify clinically relevant deep learning algorithms for emphysema quantification using low-dose chest computed tomography (LDCT) through an invitation-based competition.

MATERIALS AND METHODS

The Korean Society of Imaging Informatics in Medicine (KSIIM) organized a challenge for emphysema quantification between November 24, 2020 and January 26, 2021. Seven invited research teams participated in this challenge. In total, 558 pairs of computed tomography (CT) scans (468 pairs for the training set, and 90 pairs for the test set) from 9 hospitals were collected retrospectively or prospectively. CT acquisition followed the hospitals' protocols to reflect the real-world clinical setting. Using the training set, each team developed an algorithm that generated converted LDCT by changing the pixel values of LDCT to simulate those of standard-dose CT (SDCT). The agreement between SDCT and LDCT was evaluated using the intraclass correlation coefficient (ICC; 2-way random effects, absolute agreement, and single rater) for the percentage of low-attenuated area below -950 HU (LAA-950 HU), κ value for emphysema categorization (LAA-950 HU, <5%, 5% to 10%, and ≥10%) and cosine similarity of LAA-950 HU.

RESULTS

The mean LAA-950 HU of the test set was 14.2%±10.5% for SDCT, 25.4%±10.2% for unconverted LDCT, and 12.9%±10.4%, 11.7%±10.8%, and 12.4%±10.5% for converted LDCT (top 3 teams). The agreement between the SDCT and converted LDCT of the first-place team was 0.94 (95% confidence interval: 0.90, 0.97) for ICC, 0.71 (95% confidence interval: 0.58, 0.84) for categorical agreement, and 0.97 (interquartile range: 0.94 to 0.99) for cosine similarity.

CONCLUSIONS

Emphysema quantification with LDCT was feasible through deep learning-based CT conversion strategies.

Kernel Conversion for Robust Quantitative Measurements of Archived Chest Computed Tomography Using Deep Learning-Based Image-to-Image Translation

Chest computed tomography (CT) is used to screen for lung cancer and evaluate pulmonary and extra-pulmonary abnormalities such as emphysema and coronary artery calcification, particularly in smokers. In real-world practice, lung abnormalities are visually assessed using high-contrast thin-slice images which are generated from raw scan data using sharp reconstruction kernels with the sacrifice of increased image noise. In contrast, accurate CT quantification requires low-contrast thin-slice images with low noise, which are generated using soft reconstruction kernels. However, only sharp-kernel thin-slice images are archived in many medical facilities due to limited data storage space. This study aimed to establish deep neural network (DNN) models to convert sharp-kernel images to soft-kernel-like images with a final goal to reuse historical chest CT images for robust quantitative measurements, particularly in completed previous longitudinal studies. By using pairs of sharp-kernel (input) and soft-kernel (ground-truth) images from 30 patients with chronic obstructive pulmonary disease (COPD), DNN models were trained. Then, the accuracy of kernel conversion based on the established DNN models was evaluated using CT from independent 30 smokers with and without COPD. Consequently, differences in CT values between new images converted from sharp-kernel images using the established DNN models and ground-truth soft-kernel images were comparable with the inter-scans variability derived from repeated phantom scans (6 times), showing that the conversion error was the same level as the measurement error of the CT device. Moreover, the Dice coefficients to quantify the similarity between low attenuation voxels on given images and the ground-truth soft-kernel images were significantly higher on the DNN-converted images than the Gaussian-filtered, median-filtered, and sharp-kernel images (p < 0.001). There were good agreements in quantitative measurements of emphysema, intramuscular adipose tissue, and coronary artery calcification between the converted and the ground-truth soft-kernel images. These findings demonstrate the validity of the new DNN model for kernel conversion and the clinical applicability of soft-kernel-like images converted from archived sharp-kernel images in previous clinical studies. The presented method to evaluate the validity of the established DNN model using repeated scans of phantom could be applied to various deep learning-based image conversions for robust quantitative evaluation.

Analysis of Tracheobronchial Diverticula Based on Semantic Segmentation of CT Images via the Dual-Channel Attention Network

Tracheobronchial diverticula (TD) is a common cystic lesion that can be easily neglected; hence accurate and rapid identification is critical for later diagnosis. There is a strong need to automate this diagnostic process because traditional manual observations are time-consuming and laborious. However, most studies have only focused on the case report or listed the relationship between the disease and other physiological indicators, but a few have adopted advanced technologies such as deep learning for automated identification and diagnosis. To fill this gap, this study interpreted TD recognition as semantic segmentation and proposed a novel attention-based network for TD semantic segmentation. Since the area of TD lesion is small and similar to surrounding organs, we designed the atrous spatial pyramid pooling (ASPP) and attention mechanisms, which can efficiently complete the segmentation of TD with robust results. The proposed attention model can selectively gather features from different branches according to the amount of information they contain. Besides, to the best of our knowledge, no public research data is available yet. For efficient network training, we constructed a data set containing 218 TD and related ground truth (GT). We evaluated different models based on the proposed data set, among which the highest MIOU can reach 0.92. The experiments show that our model can outperform state-of-the-art methods, indicating that the deep learning method has great potential for TD recognition.

[Pulmonary Emphysema: Visual Interpretation and Quantitative Analysis]

Pulmonary emphysema is a cause of chronic obstructive pulmonary disease. Emphysema can be accurately diagnosed via CT. The severity of emphysema can be assessed using visual interpretation or quantitative analysis. Various studies on emphysema using deep learning have also been conducted. Although the classification of emphysema has proven clinically useful, there is a need to improve the reliability of the measurement.

New Method for Combined Quantitative Assessment of Air-Trapping and Emphysema on Chest Computed Tomography in Chronic Obstructive Pulmonary Disease: Comparison with Parametric Response Mapping

Objective

Emphysema and small-airway disease are the two major components of chronic obstructive pulmonary disease (COPD). We propose a novel method of quantitative computed tomography (CT) emphysema air-trapping composite (EAtC) mapping to assess each COPD component. We analyzed the potential use of this method for assessing lung function in patients with COPD.

Materials and Methods

A total of 584 patients with COPD underwent inspiration and expiration CTs. Using pairwise analysis of inspiration and expiration CTs with non-rigid registration, EAtC mapping classified lung parenchyma into three areas: Normal, functional air trapping (fAT), and emphysema (Emph). We defined fAT as the area with a density change of less than 60 Hounsfield units (HU) between inspiration and expiration CTs among areas with a density less than −856 HU on inspiration CT. The volume fraction of each area was compared with clinical parameters and pulmonary function tests (PFTs). The results were compared with those of parametric response mapping (PRM) analysis.

Results

The relative volumes of the EAtC classes differed according to the Global Initiative for Chronic Obstructive Lung Disease stages (p < 0.001). Each class showed moderate correlations with forced expiratory volume in 1 second (FEV₁) and FEV₁/forced vital capacity (FVC) (r = −0.659–0.674, p < 0.001). Both fAT and Emph were significant predictors of FEV₁ and FEV₁/FVC (R² = 0.352 and 0.488, respectively; p < 0.001). fAT was a significant predictor of mean forced expiratory flow between 25% and 75% and residual volume/total vital capacity (R² = 0.264 and 0.233, respectively; p < 0.001), while Emph and age were significant predictors of carbon monoxide diffusing capacity (R² = 0.303; p < 0.001). fAT showed better correlations with PFTs than with small-airway disease on PRM.

Conclusion

The proposed quantitative CT EAtC mapping provides comprehensive lung functional information on each disease component of COPD, which may serve as an imaging biomarker of lung function.

Non-invasive assessment of cirrhosis using multiphasic dual-energy CT iodine maps: correlation with model for end-stage liver disease score

PURPOSE

To determine whether multiphasic dual-energy (DE) CT iodine quantitation correlates with the severity of chronic liver disease.

METHODS

We retrospectively included 40 cirrhotic and 28 non-cirrhotic patients who underwent a multiphasic liver protocol DECT. All three phases (arterial, portal venous (PVP), and equilibrium) were performed in DE mode. Iodine (I) values (mg I/ml) were obtained by placing regions of interest in the liver, aorta, common hepatic artery, and portal vein (PV). Iodine slopes (λ) were calculated as follows: (I_equilibrium-I_arterial)/time and (I_equilibrium-I_PVP)/time. Spearman correlations between λ and MELD scores were evaluated, and the area under the curve of the receiver operating characteristic (AUROC) was calculated to distinguish cirrhotic and non-cirrhotic patients.

RESULTS

Cirrhotic and non-cirrhotic patients had significantly different λ_{equilibrium-arterial} [IQR] for the caudate (λ = 2.08 [1.39-2.98] vs 1.46 [0.76-1.93], P = 0.007), left (λ = 2.05 [1.50-2.76] vs 1.51 [0.59-1.90], P = 0.002) and right lobes (λ = 1.72 [1.12-2.50] vs 1.13 [0.41-0.43], P = 0.003) and for the PV (λ = 3.15 [2.20-5.00] vs 2.29 [0.85-2.71], P = 0.001). λ_{equilibrium-PVP} were significantly different for the right (λ = 0.11 [- 0.45-1.03] vs - 0.44 [- 0.83-0.12], P = 0.045) and left lobe (λ = 0.30 [- 0.25-0.98] vs - 0.10 [- 0.35-0.24], P = 0.001). Significant positive correlations were found between MELD scores and λ_{equilibrium-arterial} for the caudate lobe (ρ = 0.34, P = 0.004) and λ_{equilibrium-PVP} for the caudate (ρ = 0.26, P = 0.028) and right lobe (ρ = 0.33, P = 0.007). AUROC in distinguishing cirrhotic and non-cirrhotic patients were 0.72 (P = 0.002), 0.71 (P = 0.003), and 0.75 (P = 0.001) using λ_{equilibrium-arterial} for the left lobe, right lobe, and PV, respectively. The λ_{equilibrium-PVP} AUROC of the right lobe was 0.73 (P = 0.001).

CONCLUSION

Multiphasic DECT iodine quantitation over time is significantly different between cirrhotic and non-cirrhotic patients, correlates with the MELD score, and it could potentially serve as a non-invasive measure of cirrhosis and disease severity with acceptable diagnostic accuracy.