Tensorflow and vector fertility: the automatic classification of pyriproxyfen-damaged mosquito ovaries

M T Fowler, R Lees, C Ngufor, N Matowo, N Protopopoff, A. Spiers

Background & objective: Pyriproxyfen (PPF) offers an alternative to pyrethroids in areas where pyrethroid resistant vectors are prevalent. The efficacy of PPF is currently assessed through the manual inspection of vector ovary damage by human experts. However, this manual process is inefficient, inconsistent, difficult to replicate and its accuracy is hard to substantiate. Furthermore, the required expertise can be difficult to train and is not available in many contexts. Therefore, a freely available alternative method for the accurate, quick and automatic classification of ovary damage is required.

Materials & method: Using the TensorFlow library within python, a resnet-50 convolutional neural network (CNN) was pre-trained using the ImageNet dataset. This CNN architecture was then repurposed and measured using a novel dataset of 163 dissected ovary images whose fertility status and PPF exposure was known. Data augmentation was employed to maximise the training dataset and produce 2 552 random images. A test set of 47 images was used to measure accuracy.

Results & discussion: The model produced an accuracy score (correct predictions divided by total number of predictions) of 0.936 (94%) and an AUC(a comparison of true positive against false positive) of 0.902 (90%). The application of the model on the 47 images in the test set took 12.83 seconds.

Conclusion & recommendation: Reliance on experts to determine the efficacy of PPF is subject to limitations. We show that these can be overcome using a cnn model that automates the classification of ovary fertility status. Such a model can achieve an acceptable level of precision, in a quick, robust format and has the potential to be easily distributed in a practical and accessible manner. Furthermore, this approach is applicable to any ppf treated tool, or similarly acting insecticide, and is useful for measuring efficacy and in durability monitoring.