(E-200) Computational network-based virtual CT images to evaluate the performance of binary classifiers for distinguishing between stable and progressive emphysema

Introduction:: Chronic Obstructive Pulmonary Disease (COPD) is among the top four diseases claiming most deaths worldwide. Currently, there is no cure for COPD, and treatment options depend on diagnosing how fast the disease progresses. Patients that have frequent exacerbations show faster decline of lung function and accelerated evidence of emphysematous disease on lung CT images. Previously, we have shown that structural deterioration of the parenchyma is driven by mechanical forces rupturing enzymatically weakened alveolar septal walls (1). We therefore hypothesized that the faster decline of lung function in COPD exacerbators is reflected in altered tissue structure that can be identified from single CT images of the chest. To test this hypothesis, we generated synthetic CT images of the lung using an elastic network model of emphysema progression, and then used binary classifiers to identify cases corresponding to slowly versus rapidly progressing disease.


Materials and Methods:: Two-dimensional linear pre-stressed spring networks were used to represent alveolar tissue. Emphysema progression was mimicked by gradually eliminating springs from the network either randomly, mimicking a stable or slowly progressing disease (S-type), or disease based on mechanical force (F-type) that shows faster degradation representing the frequent exacerbator phenotype. After 5 steps of removing 1% of the internal springs, the networks were converted to synthetic CT images. A total of 157 different S-type and 200 different F-type networks were generated corresponding to the same number of “patients”, with each patient having 6 sequential images corresponding to progressing stages of disease. Five different binary classifiers were implemented: nearest neighbor (NN), k-nearest neighbor (kNN), linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), and a 2-layer neural network. Before classification, principal component analysis was used to optimally reduce the data dimensions from the 3721 represented in the synthetic CT images to between 50 and 300 depending on the classification method. The algorithms were trained on 80% of the data and tested on 20% of the data using a 4-fold averaging.


Results, Conclusions, and Discussions:: The progression of emphysema modeled by the network and reflected in the synthetic CT images is demonstrated in Fig. 1 (A-D). Although the same number of springs were eliminated from both networks at each step, the F-type network showed significantly more destruction. The NN and the LDA methods performed poorly whereas the kNN provided a good predictor with an average error rate of 0.7% ± 0.68%. The QDA was slightly better with error rates of 0.47% ± 0.46% whereas the neural network had an order of magnitude lower error rate of 0.023% ± 0.011%.

This study has a number of limitations. The synthetic CT images corresponding to different hypothetical subjects were relatively quite similar to each other, which likely resulted in the near zero error rate for the neural network model. Additionally, the spring network did not recapitulate the 3D structure of the lung, and the effects of gravity were also neglected. Finally, the model of emphysema progression was simplistic and did not consider the likely possibility of emphysema resulting from a mixture of random and correlated force-based tissue rupture. Despite these limitations, however, our study provides proof of concept that large numbers of synthetic CT images of the emphysematous lung method can be used for initial training of machine learning algorithms prior to eventual training refinement on images from actual patients.


Acknowledgements (Optional): :

References (Optional): : (1) Suki, B. et al. Emphysema and mechanical stress-induced lung remodeling. Physiology. 2013;28(6):404-413. doi: 10.1152/physiol.00041.2013. PubMed PMID: 24186935.