Image Processing for Diagnosis of Pulmonary Hypertension Diseases on Chest Radiographs

Introduction: Pulmonary Hypertension diseases lead to pathological changes in lung parenchyma detectableon chest X-rays. Automated image analysis may allow objective assessment. To develop image processing andmachine learning methods to detect and grade pulmonary hypertension diseases from chest X-rays.Methods: ۵۰ chest X-rays with ۲۶ hyper inflated and ۲۴ normal patients were included. Lung segmentationwas done using thresholding and active shape models. Texture features were extracted using gray level cooccurrencematrices (contrast, homogeneity), run length matrices (short run emphasis, long run emphasis), andLaws textures (level, edge, spot). Shape features like circularity and density features like kurtosis werecalculated. Feature selection using Wilcoxon rank sum test was done to determine optimal features. These wereused to train artificial neural network (ANN), support vector machine (SVM), and logistic regression modelsusing five-fold cross-validation. Binary classification for normal vs. hyper inflated lungs and multi-classclassification for mild vs. moderate vs. severe hyperinflation were evaluated.Results: Texture features based on run length matrices and Laws textures showed significant differencesbetween normal and hyper inflated lungs (p<۰.۰۱). The ANN model achieved highest accuracy of ۸۴% forbinary classification of normal vs. disease. For multi-class hyperinflation grading, the SVM model attainedoptimal accuracy of ۷۲%.Conclusion: Image processing can extract informative features from chest X-rays allowing automated diagnosisand quantification of pulmonary hypertension diseases. The machine learning models developed show potentialas decision support tools for radiologists.