Image Analysis to Automate Parasite Diagnostics

We study the development of machine learning models from microscopy images of pet stool samples to detect and classify parasite eggs. We specifically train convolutional neural networks and compare different setups regarding the type of sampling, loss functions, and types of augmentation. Our best models achieve a 93% recall in distinguishing between images containing parasite eggs and images without eggs, and the classifying model achieves 97,2% accuracy on the test set. The final goal would be to automate the process by providing a pipeline for scanning microscopy slides and operationalizing the models. We explored a few options for creating images that could be fed to the models, such as a slide scanner and a high dpi scanner.