We addressed the problem of automatic malaria diagnosis. In particular, we designed a tool for medical technicians that detects and counts the number of infected and healthy cells from blood smear images and also classifies cells into a particular life cycle of malaria infection. The tool provides a major improvement in the speed of the diagnosis process. We developed two models; MD_FCOS_c6 and MD_FCOS_c2, which both use FCOS, and model MD_FCOS_DNet_c6 which uses FCOS and DenseNet. We analyzed our models on an open-sourced dataset of blood smears which contains 1364 images. MD_FCOS_c6 and MD_FCOS_DNet_c6 models detect six classes of cells, meanwhile MD_FCOS_c2 model detects only healthy and infected cells. The MD_FCOS_c6 model achieves a F1 score above 0.7 for half of the classes and above 0.3 for the other half. MD_FCOS_c2 model, specialized in detecting the difference between infected and healthy cells, has F1 score above 0.95 for healthy cells and above 0.8 for infected cells. MD_FCOS_DNet_c6 model is better at detecting classes that are underrepresented in the database, but also worse than MD_FCOS_c6 at detecting well represented classes. MD_FCOS_DNet_c6 model recognizes an infected blood smear as infected with the probability of 98.8%. We also developed a web application that enables the usage of our model to medical technicians. The application provides a simple user interface that shows the annotated cells on the input image, providing a tool to technicians that shows a diagnosis in less than 30 seconds.
|
