Lung nodule classification with multilevel patch-based context analysis.

In this paper, we propose a novel classification method for the four types of lung nodules, i.e., well-circumscribed, vascularized, juxta-pleural, and pleural-tail, in low dose computed tomography scans. The proposed method is based on contextual analysis by combining the lung nodule and surrounding anatomical structures, and has three main stages: an adaptive patch-based division is used to construct concentric multilevel partition; then, a new feature set is designed to incorporate intensity, texture, and gradient information for image patch feature description, and then a contextual latent semantic analysis-based classifier is designed to calculate the probabilistic estimations for the relevant images. Our proposed method was evaluated on a publicly available dataset and clearly demonstrated promising classification performance.

Fully automated scoring of chest radiographs in cystic fibrosis.

We present a prototype of a fully automated scoring system for chest radiographs (CXRs) in cystic fibrosis. The system was used to analyze real, clinical CXR data, to estimate the Shwachman-Kulczycki score for the image. Images were resampled and normalized to a standard size and intensity level, then segmented with a patch-based nearest-neighbor mapping algorithm. Texture features were calculated regionally and globally, using Tamura features, local binary patterns (LBP), gray-level co-occurrence matrix and Gabor filtering. Feature selection was guided by current understanding of the disease process, in particular the reorganization and thickening of airways. Combinations of these features were used as inputs for support vector machine (SVM) learning to classify each CXR, and evaluated using two-fold cross-validation for agreement with clinician scoring. The final computed score for each image was compared with the score assigned by a physician. Using this prototype system, we analyzed 139 CXRs from an Australian pediatric cystic fibrosis registry, for which texture directionality showed greatest discriminating power. Computed scores agreed with clinician scores in 75% of cases, and up to 90% of cases in discriminating severe disease from mild disease, similar to the level of human interobserver agreement for this dataset.

Overlapping node discovery for improving classification of lung nodules.

Distinguishing malignant lung nodules from benign nodules is an important aspect of lung cancer diagnosis. In this paper, we propose an automatic method to classify lung nodules into four different types, i.e. well-circumscribed, juxta-vascular, juxta-pleural and pleural-tail. Additionally, since the morphology of lung nodules forms a continuum between the different types, our proposed method is superior to previous methods that classify single nodules into a single type. First, a weighted similarity network is constructed based on the SVM with probability estimates, turning the 128-length SIFT descriptor to a 4-length probability vector against the four types. Then, the classification of nodules while identifying those with overlapping types is made using the weighed Clique Percolation Method (CPMw). We evaluate the proposed method on low-dose CT images from ELCAP. Our results show that there is more overlap between well-circumscribed and juxta-vascular, and between juxta-pleural and pleural tail. Also, quantitative comparisons among various methods demonstrate highly effective nodule classification results by identifying the overlapping nodule types.