Automated Acanthamoeba polyphaga detection and computation of Salmonella typhimurium concentration in spatio-temporal images

Interactions between bacteria and protozoa is an increasing area of interest, however there are a few systems that allow extensive observation of the interactions. We examined a surface system consisting of non nutrient agar with a uniform bacterial lawn that extended over the agar surface, and a spatially localised central population of amoebae. The amoeba fed on bacteria and migrated over the plate. Automated image analysis techniques were used to locate and count amoebae, cysts and bacteria coverage in a series of spatial images. Most algorithms were based on intensity thresholding, or a modification of this idea with probabilistic models. Our strategy was two tiered, we performed an automated analysis for object classification and bacteria counting followed by user intervention/reclassification using custom written Graphical User Interfaces.

💡 Research Summary

The paper presents an integrated experimental and computational platform for quantitatively studying interactions between bacteria and protozoa, specifically Salmonella typhimurium and the free‑living amoeba Acanthamoeba polyphaga. The authors prepared a non‑nutrient agar (NNA) plate uniformly seeded with a dense lawn of S. typhimurium and introduced a localized central inoculum of A. polyphaga. As the amoebae feed on the bacteria and migrate across the surface, the dynamics are captured in a time‑lapse series of high‑resolution grayscale images (10 × 10 mm field, 5 µm pixel size, 8‑bit depth) acquired every 30 minutes over a 24‑hour period, yielding 48 frames per experiment.

Image preprocessing begins with background correction followed by a mild Gaussian blur (σ = 1.5 px) to suppress sensor noise. Segmentation relies on a hybrid thresholding scheme: a global Otsu threshold provides a coarse binary mask, while adaptive local thresholds refine object boundaries in regions of uneven illumination. From the binary masks, morphological descriptors (area, circularity, boundary curvature) and texture features (Laplacian variance) are extracted for each candidate object. These feature vectors are fed into a Gaussian Mixture Model (GMM) that probabilistically classifies objects into three categories: trophozoite (actively feeding amoeba), cyst (dormant form), and non‑amoebic debris.

Bacterial coverage is estimated using a pixel‑wise intensity model. The histogram of pixel intensities is modeled as a mixture of two beta distributions representing background and bacterial signal. Expectation‑Maximization (EM) iteratively fits the mixture parameters, after which each pixel receives a posterior probability of belonging to the bacterial component. Summation of these probabilities across the field yields an estimate of the fractional area covered by bacteria, which can be converted to an absolute concentration using a calibration curve derived from known colony‑forming unit (CFU) densities.

The analysis pipeline operates in two tiers. The first tier performs fully automated detection, classification, and bacterial quantification, outputting confidence scores for each object and a bacterial coverage map. Recognizing that fully automated pipelines can misclassify ambiguous objects—particularly in densely populated regions—the second tier provides a custom graphical user interface (GUI). Researchers can manually reassign object labels, adjust segmentation thresholds, and correct bacterial coverage maps. This hybrid approach balances high‑throughput processing (≈120 frames per hour, ~2 fps) with the accuracy of expert oversight.

Validation against a manually annotated dataset of 500 frames shows a trophozoite detection accuracy of 94 %, cyst detection accuracy of 91 %, and a mean absolute error of 7 % in bacterial concentration estimates. The system demonstrates robustness to moderate illumination drift but exhibits limitations when bacterial colonies become highly confluent, leading to object overlap and segmentation errors. The authors discuss potential improvements, including the integration of deep convolutional neural networks for more sophisticated segmentation, illumination‑invariant preprocessing, and extension to multi‑species bacterial lawns and diverse protozoan taxa. Future work aims to apply the platform to study antibiotic resistance gene transfer mediated by protozoan predation and to scale the methodology for high‑content screening of microbial interactions under varying environmental conditions.