Skip to main content

Research Repository

Advanced Search

Detection of morphological changes caused by chemical stress in the cyanobacterium Planktothrix agardhii using convolutional neural networks. [Dataset]

Contributors

Ismael Carloto
Data Collector

Abstract

The presence of harmful algal bloom in many reservoirs around the world, alongside the lack of sanitation law/ordinance regarding cyanotoxin monitoring (particularly in developing countries), create a scenario in which the local population could potentially chronically consume cyanotoxin-contaminated waters. Therefore, it is crucial to develop low cost tools to detect possible systems failures and consequent toxin release inferred by morphological changes of cyanobacteria in the raw water. This paper aimed to look for the best combination of convolutional neural network (CNN), optimizer and image segmentation technique to differentiate P. agardhii trichomes before and after chemical stress caused by the addition of hydrogen peroxide. The file accompanying this output contains supplementary figures and tables of results obtained from the experiments in this study.

Citation

CARLOTO, I., JOHNSTON, P., PESTANA, C.J. and LAWTON, L.A. 2021. Detection of morphological changes caused by chemical stress in the cyanobacterium Planktothrix agardhii using convolutional neural networks. [Dataset]. Science of the total environment [online], 784, article 146956. Available from: https://www.sciencedirect.com/science/article/pii/S004896972102026X#s0105

Acceptance Date Mar 31, 2021
Online Publication Date Apr 7, 2021
Publication Date Aug 25, 2021
Deposit Date Apr 26, 2021
Publicly Available Date Apr 8, 2022
Publisher Elsevier
DOI https://doi.org/10.1016/j.scitotenv.2021.146956
Keywords Cyanobacteria; CNN; Water treatment; Hydrogen peroxide; Image recognition; Image segmentation
Public URL https://rgu-repository.worktribe.com/output/1323443
Related Public URLs https://rgu-repository.worktribe.com/output/1254826
Type of Data Supplementary 13 tables and 14 figures.
Collection Date Mar 31, 2021
Collection Method Axenic Planktothrix agardhii CCNP 1305 culture was incubated in BG-11 (Rippka et al., 1979) medium at 20 ± 1 °C with 12 h/12 h light/dark cycle under cool white fluorescent light with an intensity of 20 μmol photons m−2 s−1. Cell density was estimated using an Olympus microscope (Model BX53M), with a Sedgewick-Rafter chamber at 200× magnification. In order to train a CNN in recognising the morphological changes in P. agardhii CCNP 1305 it was necessary both to choose a chemical compound and determine its ideal concentration that induces morphological changes rather than complete destruction of the organisms. To find the optimum H2O2 concentration three batch-type experiments were performed. Hydrogen peroxide was added to P. agardhii CCNP 1305 cell suspension (SM Table S1) to achieve final concentrations of 5, 10, 15, 20 mg L−1 in 40 mL of H2O2. A second experiment was carried out with three different concentrations of 30, 40 and 80 mg L−1 H2O2. Finally, a third, confirmatory, study was performed repeating concentrations 40 and 80 mg L−1 H2O2. In all experiments P. agardhii CCNP 1305 cultures were exposed to H2O2 with a contact time of 24 h at 20 ± 1 °C and constant cool white fluorescent light with 40 μmol photons m−2 s−1 intensity. Samples were taken before the addition of H2O2 and after 0.5, 1, 3, 6 and 24 h. Flask for each concentration were prepared in triplicate. The RGB (Red, Green, Blue) images were captured using a YenCam HD (Yenway Microscopes) camera at 500× magnification and a resolution of 1920 × 1080 pixels. Different points of the microscope slides were randomly selected to capture the images, in order to ensure that no P. agardhii trichome was captured more than once. The CNN models were implemented using Keras/TensorFlow python libraries, on an Omen Hp laptop with 16GB RAM (HP Inc., USA), Intel Core i7 2.6 GHz central processing unit and equipped with a NVIDIA GeForce RTX 2060 graphics card. In order to determine a suitable CNN architecture three different architectures were tested: one simple architecture loosely based on LeNet5 (LeCun et al., 1998), from now on called 3ConvLayer, a variation on AlexNet (Krizhevsky et al., 2012) and one based on a Keras implementation which is known to achieve 99.25% accuracy on MNIST (LeCun et al., 1998), from now on called 2ConvLayer. Three different CNN architectures with four different optimizers on image datasets that were pre-processed in six different ways were tested. Randomisation was inherent in the initial configuration of our CNNs as was allocation of images to test/train sets. A more detailed description of the Materials and Methods used can be found in section 2 of the published article (https://doi.org/10.1016/j.scitotenv.2021.146956 ).