Latent traits of lung tissue patterns in former smokers derived by dual channel deep learning in computed tomography images
Frank LiJiwoong ChoiChunrui ZouJohn D. NewellAlejandro P. ComellasChang Hyun LeeHongseok KoR. Graham BarrEugene R. BleeckerChristopher B. CooperFereidoun AbtinIgor BarjaktarevićDavid CouperMeiLan K. HanNadia N. HanselRichard E. KannerRobert PaineElla A. KazerooniFernando J. MartínezWanda K. O’NealStephen I. RennardBenjamin M. SmithPrescott G. WoodruffEric A. HoffmanChing-Long Lin
16
Citation
55
Reference
10
Related Paper
Citation Trend
Abstract:
Abstract Chronic obstructive pulmonary disease (COPD) is a heterogeneous disease and the traditional variables extracted from computed tomography (CT) images may not be sufficient to describe all the topological features of lung tissues in COPD patients. We employed an unsupervised three-dimensional (3D) convolutional autoencoder (CAE)-feature constructor (FC) deep learning network to learn from CT data and derive tissue pattern-clusters jointly. We then applied exploratory factor analysis (EFA) to discover the unobserved latent traits (factors) among pattern-clusters. CT images at total lung capacity (TLC) and residual volume (RV) of 541 former smokers and 59 healthy non-smokers from the cohort of the SubPopulations and Intermediate Outcome Measures in the COPD Study (SPIROMICS) were analyzed. TLC and RV images were registered to calculate the Jacobian (determinant) values for all the voxels in TLC images. 3D Regions of interest (ROIs) with two data channels of CT intensity and Jacobian value were randomly extracted from training images and were fed to the 3D CAE-FC model. 80 pattern-clusters and 7 factors were identified. Factor scores computed for individual subjects were able to predict spirometry-measured pulmonary functions. Two factors which correlated with various emphysema subtypes, parametric response mapping (PRM) metrics, airway variants, and airway tree to lung volume ratio were discriminants of patients across all severity stages. Our findings suggest the potential of developing factor-based surrogate markers for new COPD phenotypes.Keywords:
Autoencoder
Stacked autoencoder is a typical deep neural network. The hidden layers will compress the input data with a better representation than the raw data. Stacked autoencoder has several hidden layers. However, the number of hidden layers is always experiential. In this paper, different hidden layers number autoencoders are discussed. Different depths of stacked autoencoder have different learning capability. The deeper stacked autoencoders have better learning capability which needs more training iterations and time.
Autoencoder
Feature Learning
Representation
Cite
Citations (20)
This paper presents a comparison performance on three types of autoencoders, namely, the traditional autoencoder with Restricted Boltzmann Machine (RBM), the stacked autoencoder without RBM and the stacked autoencoder with RBM. The performances are compared based on the reconstruction error for face images and using the same values for the parameters such as the number of neurons in the hidden layers, the training method, and the learning rate. The results show that the RBM stacked autoencoder gives better performance in terms of the reconstruction error compared to the other two architectures.
Autoencoder
Restricted Boltzmann machine
Word error rate
Cite
Citations (28)
Autoencoder is an excellent unsupervised learning algorithm. However, it can not generate kinds of sample data in the decoding process. Variational autoencoder is a typical generative adversarial net which can generate various data to augment the sample data. In this paper, we want to do some research about the information learning in hidden layer. In the simulation, we compare the hidden layer learning of hidden layer in conventional autoencoder and variational autoencoder.
Autoencoder
Generative model
Sample (material)
Cite
Citations (22)
The anomaly detection technology is the basis for ensuring the safe and stable operation of the on-rail payload. The traditional threshold-based anomaly detection method has low accuracy and poor flexibility, and cannot detect abnormalities in real time. In addition, due to the lack of abnormal samples, the distribution of positive and negative samples is extremely imbalanced, which increases the difficulty of abnormal detection. Therefore, this paper proposes an unsupervised learning method based on AutoEncoder and its variants, the Basic AutoEncoder, Deep AutoEncoder and Sparse AutoEncoder are used to verify the algorithm on three public datasets. And using the above three algorithms to carry out the case application on the real load dataset. The experiments show whether in the public dataset or the real data of the payload, the three methods of AutoEncoder have achieved good results, proving the AutoEncoder and its variants have a good application in anomaly detection. At the same time, it is verified that the three algorithms have different effects on different datasets, which proves that the AutoEncoder with different characteristics need to be selected in different scenarios.
Autoencoder
Payload (computing)
Anomaly (physics)
Cite
Citations (0)
Deep Autoencoder has the powerful ability to learn features from large number of unlabeled samples and a small number of labeled samples. In this work, we have improved the network structure of the general deep autoencoder and applied it to the disease auxiliary diagnosis. We have achieved a network by entering the specific indicators and predicting whether suffering from liver disease, the network using real physical examination data for training and verification. Compared with the traditional semi-supervised machine learning algorithm, deep autoencoder will get higher accuracy.
Autoencoder
Training set
Cite
Citations (0)
Autoencoder
Transfer of learning
Cite
Citations (2)
Background
COPD is characterised by reduced airway lumen dimensions and fewer peripheral airways. Most studies of airway properties sample airways based upon lumen dimension or at random, which may bias comparisons given reduced airway lumen dimensions and number in COPD. We sought to compare central airway wall dimensions on CT in COPD and controls using spatially matched airways, thereby avoiding selection bias of airways in the lung.Methods
The Multi-Ethnic Study of Atherosclerosis (MESA) COPD Study and Subpopulations and Intermediate Outcomes in COPD Study (SPIROMICS) recruited smokers with COPD and controls aged 50–79 years and 40–80 years, respectively. COPD was defined by current guidelines. Using CT image data, airway dimensions were measured for all central airway segments (generations 0–6) following 5 standardised paths into the lungs. Case-control airway comparisons were spatially matched by generation and adjusted for demographics, body size, smoking, CT dose, per cent emphysema, airway length and lung volume.Results
Among 311 MESA COPD participants, airway wall areas at generations 3–6 were smaller in COPD compared with controls (all p<0.001). Among 1248 SPIROMICS participants, airway wall areas at generations 1–6 were smaller (all p<0.001), and this reduction was monotonic with increasing COPD severity (p<0.001). In both studies, sampling airways by lumen diameter or randomly resulted in a comparison of more proximal airways in COPD to more peripheral airways in controls (p<0.001) resulting in the appearance of thicker walls in COPD (p<0.02).Conclusions
Airway walls are thinner in COPD when comparing spatially matched central airways. Other approaches to airway sampling result in comparisons of more proximal to more distal airways and potentially biased assessment of airway properties in COPD.Lumen (anatomy)
Airway obstruction
Cite
Citations (136)
BackgroundDiagnosing Chronic Obstructive Pulmonary Disease (COPD) without spirometry is difficult; we had developed previously a scale (DS-COPD) in an epidemiological setting. It allowed diagnosing COPD confidently when scored high, and excluded confidently when low.AimTo validate the DS-COPD in clinical setting through a case-control study, and to evaluate the cost saving by its use.MethodsIn two tertiary care hospitals, we calculated the DS-COPD scale in suspected COPD and controls; COPD was predicted in the study sample and in symptomatic individuals. COPD status was confirmed by post-bronchodilator spirometry.ResultsFrom the ROC curve, the Area Under Curve was 0.945. The Positive Predictive Value was 79% if DS-COPD was >17 and the Negative Predictive Value was 83% if DS-COPD was <10 in symptomatic individuals. A DS-COPD of 10–17 represented a gray zone mostly suggestive of no COPD. For every 100 symptomatic patients 4150$ were saved combining spirometry and scale when inconclusive compared to systematic use of spirometry.ConclusionsWe were able to validate a scale (DS-COPD) for COPD diagnosis in clinical setting. It would be valuable in primary care settings, where spirometry may not be available and in clinical settings before availability of spirometry results. Future prospective studies are still needed to confirm its value.
Cite
Citations (4)
Gold standard (test)
Hyperinflation
Cite
Citations (50)
A method for explaining a deep learning model prediction is proposed. It uses a combination of the standard autoencoder and the variational autoencoder. The standard autoencoder is exploited to reconstruct original images and to produce hidden representation vectors. The variational autoencoder is trained to transform the deep learning model outputs (embedding vectors) into the hidden representation vectors of the standard autoencoder. In explaining or testing phase, the variational autoencoder produces a set of vectors based on the explained image embedding. Then the trained decoder part of the standard autoencoder reconstructs a set of images which form a heatmap explaining the original explained image. In fact, the variational autoencoder plays a role of the perturbation technique of images. Numerical experiments with the well-known datasets MNIST and CIFAR10 illustrate the propose method.
Autoencoder
MNIST database
Representation
Cite
Citations (5)