Development of a machine learning model for malaria parasite detection in peripheral blood smears: a preliminary evaluation using Create ML

Background Microscopic examination of peripheral blood smears (PBSs) remains the gold standard for malaria diagnosis but is time-consuming and prone to interobserver variability. Artificial intelligence-based approaches can assist in malaria parasite (MPs) detection; however, many existing models require coding expertise or lack integration into laboratory workflows. The study aims to demonstrate the development and deployment of a code-free machine learning (ML) model for MP detection using a benchtop digital laboratory microscope. Methods In this retrospective study, archived Leishman-stained PBS slides positive for Plasmodium vivax and P. falciparum were digitised at 1000× magnification. 422 fields across 40 slides were digitised, segmented and augmented to yield 7041 images, which were divided into training, validation and test subsets in an 80:10:10 ratio. Ring forms, trophozoites, leucocytes and platelets were annotated in the images. Apple’s Create ML software platform was used for model training. The outcomes were assessed using intersection over union at a 50% threshold (IoU 50 ), F1-score (harmonic mean of precision and recall) and confusion matrix analysis. The trained model was deployed on a digital benchtop laboratory microscope for live object detection overlay. Results The ML model demonstrated a localisation accuracy (IoU 50 ) of 89% and 88% during training and validation, respectively. During testing, a mean IoU 50 of 80% was obtained, ranging from 64% for ring forms to 95% for trophozoites, with a mean precision of 92%, a recall of 93% and an F1 score of 0.93 for object classification. Conclusions This study demonstrates the feasibility of developing an ML model for MP detection using a code-free platform and implementation on a digital laboratory microscope.