A novel approach for testing water safety using deep learning inference of microscopic images of unincubated water samples

Fecal-contaminated water causes diseases and even death. Current microbial water safety tests require pathogen incubation, taking 24–72 h and costing $20-$50 per test. This paper presents a solution (DeepScope) that classifies microscopic images of unincubated water samples by a composite water-safety determination, meeting UNICEF’s ideal Target Product Profile numerical performance thresholds for presence/absence testing, with an estimated per-test cost of $0.44. By eliminating the need for pathogen incubation, DeepScope reduces testing time by over 98%. In DeepScope, a dataset of microscope images of bacteria and water samples was assembled from the public DIBaS bacterial image collection and water samples from multiple source types, with labels cross-validated using color-change coliform tests. A novel tile-swap augmentation technique was developed to expand dataset diversity and mitigate overfitting, swapping random tile pairs to preserve bacterial morphology while producing a large space of augmented images. Four convolutional neural network models (ResNet-18 and ResNet-50 backbones, with and without dropout regularization) were developed using transfer learning, then evaluated on a field-test dataset comprising 100,000 microscope images of unseen, real-world water samples collected from fourteen different water sources across Sammamish, WA. On the 160 original microscope acquisitions underlying this dataset, the DeepScope model (ResNet-50 with dropout) achieved 96.88% accuracy, 98.61% recall, and 94.67% precision, corresponding to a 1.39% false-negative rate and a 4.55% false-positive rate. A larger field test was then carried out on the 100,000 augmented images derived from these acquisitions, with the operating threshold set to fix precision at 90%. Under these conditions the model achieved recall of 94.48% ±0.11%, accuracy of 92.79% ±0.08%, and F1 of 92.19% ±0.09%—a 5.52% false-negative rate, well below UNICEF’s 10% ceiling and competitive with membrane filtration and color-change coliform tests without requiring pathogen incubation. The DeepScope model was tested in a web server environment using mobile applications for Android and iOS. The testing validated Internet-based or smartphone-based water safety testing, with results obtained in seconds.