Three Stage Method Assisting Endobronchial Tumor Mass Recognition in Bronchoscopy Images

Current problem of computer assistance of bronchoscopy neoplasmatic lesion interpretation is still open and exceptionally challenging question. Strongly diversified pathology appearance, lack of reference patterns, subjective criteria of lesion assessment and often unsatisfactory quality of registered video limit diagnostic effectiveness of bronchoscopic studies. Fundamental part of presented research is an analysis of various endobronchial tumor manifestations in a context of acquired image sequences, bronchoscope navigation, artifacts, lightening and reflections, changing color dominants, unstable focus conditions, etc. Representative patterns of pathologies in bronchoscopy images and standardized features of malignancy were sought. Proposed method of neoplasmatic areas indication was based on three fundamental steps of preprocessed video analysis: 1) informative frame selection, 2) block-based unsupervised determining of enlarged textural activity, 3) indication of the regions with potentially tumor signatures, based on feature selection in different domains of transformed image followed by SVM classification. Efficiency of pathology recognition was verified with a reference image dataset of 14 examinations containing diversified neoplasmatic endobronchial tumor patterns. Experimental results reveal promising accuracy of recognition for differential forms of endobronchial tumor manifestations. Test set of selected informative blocks was classified as pathological or normal with accuracy of 85%.