Classification of bacterial morphotypes from images of ZN-stained sputum-smears towards diagnosing drug-resistant TB

We describe a method for identifying and classifying acid-fast bacilli (AFB) and their associated morphotypes in the microscope-images of Ziehl-Neelsen stained sputum smears, in the context of tuberculosis (TB) screening by image processing. The importance of our work stems from the fact that the transformation of the classical rod-shaped AFB into certain other shapes is said to be related to TB drug-resistance. The first stage of processing involves color-segmentation in the HSV space by using Neural Networks and RUS-Boosted Decision Trees. The latter is used to alleviate the effects of class-imbalance between the pixels belonging to the AFB and the background. The second stage involves categorizing the bacilli into regular rod-shaped ones (possibly beaded), their morphotypes (“V-shaped” or “Y-shaped” bacilli), and clumps. The main, and novel contribution in this paper involves identifying and classifying the bacterial morphotypes. For that purpose, we propose and investigate three different methods: The first involves assuming the morphotypes to be letters of the English alphabet, and using a letter-recognition technique based on the Hotelling Transform and the Discrete Cosine Transform on the color-segmented bacilli. The second method uses moment-based invariants on the silhouettes, boundaries and skeletons, respectively. We use Support Vector Machine and Weighted K-NN classifiers in both the cases. In addition, we describe a new method based on the ends of the skeleton. Experiments on 72 images of sputum-smears revealed that the skeleton-based approach performed better than the other methods.