Automated Enumeration and Classification of Bacteria in Fluorescent Microscopy Imagery

We present a system of techniques for automatic segmentation, quantification, and morphotype classification of vaginal bacteria from multi-band fluorescent microscopic imagery. Individual bacteria segmentation is accomplished via data pre-processing, blobness enhancement, thresholding, and multi-scale morphological decomposition. A new spotness feature is devised and extracted to effectively quantify bacterial morphotypes. A supervised classifier is trained on microscopic scans containing thousands of bacteria. Our approach is able to predict and segment bacteria with a high accuracy. The average classification error in terms of bacteria composition ratio is 6% relative to the ground-truth.