Detection and Identification of Parasite from Micro graph

Abstract :

Because of the progress of medical and sanitary environment, the number of cases of diseased due to parasite is decreasing. However, a certain number of diseases due to parasite are still reported every year. Lower frequency of occurrence of parasite related diseases, it may not easy to identify. When a small suspicious object is found from micrograph in the process of diagnosis, above mentioned circumstances may prevent the object from being identified fast. In this talk, our recent result of parasite detection and identification for micrograph is presented.

Deep neural network based methods as well as local feature based feature representation cooperated with bag of feature model contributed to achieve significant improvement of accuracy for the task of detection/classification of general object, such as human, animals, vehicles. However, there are much less distinctive visual features in genera of parasite, especially protozoa. It makes detection and identification task with acceptable accuracy more difficult. We have designed our own DNN based on RetinaNet in order to segment the region of protozoa in a micrograph, which may contain other objects specific to samples taken from human body. After the process of segmentation, local features which are found in the region of protozoa are classified based on the range of scale. Finally, identification of the species is carried out based on the similarity of statistical distribution of local features based on classes of scales.

(a) Giardia Lamblia

(b) Sarcocystis

Figure: Examples of segmentation of protozoa. Left half of each example corresponds to original micrograph with detected bounding boxes. Right half shows the result of segmentation colored with green (or brightest region in b/w printing).

*This work is supported by JSPS KAKENHI Grant NumberJP17K20025.

Keywords: Local Feature Clustering, DCNN, Parasite, Protozoa.