Automated lung cancer detection by the analysis of glandular cells in sputum cytology images using scale space features

Abstract

Out of all various types of lung cancers, adenocarcinoma is increasing at an alarming rate mainly due to the increased rate of smoking. This work aims at developing a sputum cytology image analysis system which identifies benign and malignant glandular cells. In our proposed system, we develop an automated lung cancer detection system which segments the cell nuclei and classifies the glandular cells from the given sputum cytology image using a novel scale space catastrophe histogram (SSCH) feature. Catastrophe points occur when pairwise annihilation of extrema and saddle happens in scale space. Unusual nuclear texture shows the presence of malignancy in cells, and SSCH-based texture feature extraction from nuclear region is done. From the input high-resolution image, the cellular regions are localized using maximization of determinant of Hessian, nuclei regions are segmented using K-means clustering, and SSCH features are extracted and classified using support vector machine and color thresholding. The experimental results show that the proposed method obtained an accuracy of 87.53 % which is better than Gabor filter-based gray-level co-occurrence features, local binary pattern, and complex Daubechies wavelet-based features. The results obtained are in accordance with the dataset classified by medical experts.