Computer-aided Detection of Acid-fast Bacilli in Digital Pathology

During the past decades, pathology has proved to be an accurate tool for diagnosing tuberculosis. Traditional microscopic screening often takes several ten minutes for one slide while whole slide digitalization only takes about several minutes, which makes storing, remote diagnosis and mass screening possible. However, manual screening of whole slide scanning image is still time-consuming. In this study, a computer-aided (CADe) system for bacilli detection was proposed to accelerate the screening process. Color information was used to segment the bacilli. The likelihoods of being a bacilli of the regions were estimated using the quantitative morphology, color and texture information. Logistic regression was further used for false positive reduction. A dataset of 162 image blocks collected from 4 slide samples were used in the experiment. As a result, the proposed system achieved sensitivities of 100%, 90%, 80%, 70% and 60% with FP/block of 16.35, 1.96, 0.86, 0.51 and 0.36, respectively. The figure of merit (FOM) of the combination of three feature sets is 0.87 which significantly outperforms other feature sets. Summarily, the experiment results support our proposed CADe bacilli detection system to be applied in the clinical use.