2024 Volume 47 Issue 5 Pages 139-150
Although there are examples of research aiming to automate the microscopy inspection conducted in the maintenance of activated sludge treatment, the challenge lies in the necessity to collect a large number of images of various microfauna species as training data for model construction. In this study, we aimed to develop an image analysis model that eliminates the need for the annotation of microfauna species and conducts comprehensive detection and classification of such species based on their appearances. Specifically, we constructed a deep learning model composed of a microfauna detector, a feature extractor, and a microfauna classifier, utilizing knowledge of object detection and self-supervised learning. We then quantitatively evaluated the accuracies of the detection and classification of microfauna species and the relationship between the microfauna community composition obtained using this model and the bacterial community composition in activated sludge flocs determined by 16S rDNA amplicon analysis. The microfauna detector achieved AP50 = 82.85% performance, and the microfauna classifier, using supervised learning, achieved 83.3% accuracy, excluding half of the 12 false positive images containing non-microfauna objects. Furthermore, we found a significant correlation between the microfauna community composition obtained using the model applied to 12 activated sludge samples and the bacterial community composition obtained by the rDNA analysis.
