Bacilli Segmentation with Hybrid Classifier for Tuberculosis Severity Detection using Sputum Images

This study introduces a novel methodology for sputum image-based tuberculosis severity detection. The first step in this work is to improve the standard of the input sputum image by using the median filtering approach. Then the preprocessed image is segmented effectively, separating the overlapping objects by using the Improved Deep Contour Aware Network. For improving the tuberculosis severity detection process, extract specific features like Pyramid Histogram of Oriented Gradients (PHOG), Improved Median Ternary Pattern (IMTP), Local Arc Pattern, and statistical features extracted from the segmented image. In this, the IMTP feature captures the texture information effectively based on the normalized median value of pixel intensities. Then, combined all the features to form the feature set, which is subjected to the detection process. A hybrid classification model is proposed for Tuberculosis severity detection, which is the combination of the Improved LinkNet model (ILNT) and the Deep Convolutional Neural Network (DCNN) model. Initially, classify the disease using the ILNT + DCNN model based on the extracted features to determine whether bacilli are present or absent. Then, to assess the severity of the disease, compute the bacilli count by analyzing the density ratio to calculate the disease severity. The ILNT + DCNN acquired the greatest accuracy, precision, sensitivity, and MCC of 0.949, 0.947, 0.950 and 0.897, respectively.