Deep learning-based identification of Plasmodium vivax and Plasmodium falciparum species and life-cycle stages from thin blood smear images

Aims As malaria elimination advances, low parasitaemia infections and declining microscopy expertise have made identification of Plasmodium species challenging. Although microscopy remains the diagnostic gold standard, it is prone to human error and is time consuming. Deep learning algorithms offer a supportive diagnostic framework to enhance rapid, automated, and accurate detection of malaria parasites in peripheral blood smear images. Methods and Results We collected peripheral blood samples from 35 malaria-positive patients confirmed by microscopy, rapid diagnostic test and nested PCR. A total of 617 microscopy images were annotated for normal blood components and various life-cycle stages of P. vivax and P. falciparum, creating an 11-class dataset for training the YOLOv8n model. An additional 330 same-slide hold-out image set was used for evaluation. The dataset comprised 16 675 annotated cell images of infected and uninfected blood cells. The YOLOv8n model achieved a mean average precision (mAP @ 0.5) of 0.673, with mAP @ 50 greater than 0.85 for P. vivax schizonts and gametocytes, P. falciparum trophozoites and gametocytes, and white blood cells (WBCs), and an overall accuracy of 99% for distinguishing infected from uninfected cells. Sensitivity was highest for P. falciparum gametocytes (100%), with F1-scores > 0.90 for eight of eleven classes, however, the performance varied across life-cycle stages. Evaluation on the same-slide-hold-out image set demonstrated moderate-to-good performance across major parasite stages. The model also achieved high accuracy for detection of WBCs (98.1%) and platelets (99.3%). Conclusions This study presents a proof-of-concept for a YOLOv8n-based deep learning framework for species- and stage-level malaria detection from blood smears.