Quantification of malaria parasitaemia using trainable semantic segmentation and capsnet

Abstract

Malaria is a life-threatening mosquito (Anopheles)-borne blood disease caused by the plasmodium parasite. Microscopic examination of peripheral blood smears by experts helps to identify parasites precisely. The manual assessment technique is a tedious and time-consuming process. The present study focuses on developing a hybrid screening algorithm for automated identification and classification of malaria parasite-infected red blood cells (RBCs). Initially, a semantic blood cell segmentation method is adopted where a supervised classifier-regulated pixel-based segmentation is adopted to segment individual RBC present in an image. In pixel-based classification, foreground (RBCs) and background regions are considered, a pixel-based large feature dataset is generated, and an artificial neural network (ANN) classifier is trained. The trained model generates a probability map of an image which is later post-processed by Graph-cut and Marker-controlled Watershed method for developing cropped RBC image set. The proposed segmentation method achieves 99.1% accuracy. Finally, a trained modified Capsule Network (CapsNet) model is used for classification of segmented blood cells to identify the species and stages of the parasites. Here, two specific parasite species viz., Plasmodium vivax and Plasmodium falciparum with stages are considered for classification. The performance of the proposed two-steps hybrid malaria screening is promising and the training and testing on local and benchmark dataset with respect to ground truth yield 98.7% accuracy.