Separation of overlapping bacilli in microscopic digital TB images

The sputum smear microscopy based tuberculosis (TB) screening method is a conventional method employed for disease identification. It provides significant benefit to TB burdened communities across the globe; however, there are many challenges faced in processing the sputum smear images. When the smear is thick or uneven the number of overlapping bacilli is more which impedes the diagnosis. The separation of overlapping bacilli is significant without which the results lead to gross errors in identification of the disease causing agent. In this work, separation of overlapping bacilli is carried out by method of concavity (MOC) and is compared with the conventional methods such as multi-phase active contour (MAC) and marker-controlled watershed (MCW). Performance of the methods is evaluated based on the statistical mean quality score of shape descriptors extracted from the separated and existing true bacilli. The shape descriptors employed in this work include geometric features, Hu's, Zernike moments and Fourier descriptors. Results of separated overlapping bacilli demonstrate that MOC performs better than MAC and MCW. It is observed that the statistical mean quality score of the separated bacilli using the proposed MOC shows nearest match with true bacilli. The validation performed with experimental results to that of human annotations highlights the performance of MOC in separating the overlapping bacilli in the sputum smear images.