An Automated Malaria Cells Detection from Thin Blood Smear Images using Yolov4

Abstract

Malaria is a severe global health problem, with an estimated 241 million malaria infections and 627,000 malaria deaths globally in 2020. Hundreds of millions of blood films are examined annually for malaria, which includes manually counting parasites and infected red blood cells by a trained microscopist. Segmented red blood cells play an important role in applying deep learning for malaria diagnosis. However, traditional segmentation and separation of single red blood cells is challenging and requires much human intervention. Therefore, instead of segmented red blood cells, the performance of deep learning models can be studied using bounded cell images. Various object detection architectures are studied in detecting red blood cells from thin blood smear images. However, there is a lack of study on the performance of Yolov4 to detect infected cells in thin blood smear images. This study aims to evaluate the performance of Yolov4 in detecting red blood cells infected by four types of malaria species and integrate a separate algorithm to automatically crop the infected cells. Different types of malaria images are used to study if the model can still detect cells infected by various malaria parasites and from multiple stages of infection despite their morphology differences. The MP-IDB malaria datasets were used in the experiments. The performance of the Yolov4 model was evaluated by partitioning the train and test dataset by 90/10 and 80/20. The partitioning was done on datasets with and without augmentations. The results show that upon training Yolov4 model can detect infected cells despite their morphological differences. Model 4 with 80/20 dataset partition and augmentation is chosen as the best model with the best mAP of 93.43%.