Microscopic Sample Segmentation by Fully Convolutional Network for Parasite Detection

Patryk Najgebauer, Rafal Grycuk, Leszek Rutkowski, Rafal Scherer, Agnieszka Siwocha

Abstract: This paper describes a method of pixel-level segmentation applied to parasite detection. Parasite diseases in most cases are detected by microscopic samples examination or by ELISA blood tests. The microscopic methods are less invasive and often used in veterinary, but they need more time to prepare and visually evaluate samples. Diagnosticians search the entire sample to find parasite eggs and to classify their species. Depending on the species of the diagnosed animal, the samples can contain various types of pollution, e.g. fragments of plants. Most of the objects in the sample by their transparency look similar, and some of parasites eggs might be unintentionally omitted. The presented method based on fully convolutional network allows processing the entire space of the sample and assigning a class to each pixel of the image. Our model was trained to classify parasite eggs and distinguish them from adjacent or overlapped pollution.