Development of Computer Aided Diagnosis Methods for Efficient Tuberculosis Detection

Abstract

We have developed CAD methods for better detection of TB using both digital CXRs and images of ZN-stained smear microscopy. CAD methods help in providing second opinion to CXR readers, and their implementation is likely to enhance the performance of radiologists in TB diagnosis. Gist, Gabor, HOG, and PHOG feature extraction methods enabled the SVM classification model to discriminate between the digital CXR images of TB and non-TB with high accuracy. The validations of performance of these features using independent datasets and 5-fold cross-validation shows that they perform significantly better textural features currently used. Most of the earlier CAD methods were mainly based on the extraction of GLCM textural features from the segmented digital CXR images., PHOG and Gist features provided best 5-fold cross-validation accuracy for the classification of CXRs as TB or Normal for both DA and DB. These features were used to develop TB-Xpredict, which can be used for training and classification of digital CXRs. newlineZN-stained sputum smear microscopy is an essential step in the diagnosis of TB according to the guidelines of the WHO. Automation for grading of bacilli is highly desired to reduce human errors as well as mass screening of smear slides. Automation requires execution of loading of slides, focusing, stitching and counting of bacilli in an automated manner. The current work outlined and implemented a method for automated stitching of view fields. The SIFT features were used for feature detection and extraction, and RANSAC was used for removing outlier. The divide and conquer algorithm was implemented to reduce the time requirement for stitching. This method not only decreased the time complexity of the method but also provided significantly better stitching results as compared to linear stitching. The validation was done through similarity measures between the stitched and the original image both visually and quantitatively.