Development of a Deep Learning Model for Malaria Parasite Detection from Giemsa-Stained Thick Blood Films in Nigerian Rural Laboratories

Accurate malaria diagnosis by microscopy of Giemsa-stained thick blood films is the gold standard for malaria case management, but skilled microscopist availability in rural Nigerian laboratories is critically limited, motivating the development of automated image analysis tools to support microscopy-based diagnosis. This study develops and evaluates a deep learning model for automated malaria parasite detection and quantification from Giemsa-stained thick blood film images acquired under rural laboratory microscope conditions in North East Nigeria. A dataset of three thousand digital microscope images was collected from malaria-positive and negative patient blood films at health facilities in Borno and Gombe states using a smartphone microscope adapter on standard laboratory microscopes available at the field sites. Image labelling was performed by two expert malaria microscopists using a custom annotation tool, with disagreements resolved by consensus. A YOLOv5 object detection architecture was trained to detect and classify individual trophozoites, ring forms, and gametocytes of Plasmodium falciparum from thick film images, with augmentation including brightness variation, rotation, and JPEG quality degradation to simulate field imaging conditions. Model evaluation on a held-out test set of six hundred images showed precision of 88.4 percent, recall of 86.7 percent, and F1 score of 0.876 for ring-stage trophozoite detection. Parasite density estimation from model-counted parasites correlated with expert microscopist counts at Pearson r of 0.91. Model inference time per image was 0.3 seconds on a mid-range Android smartphone processor. The study demonstrates the feasibility of smartphone-based automated thick film malaria diagnosis for rural laboratory application. Keywords: malaria diagnosis, deep learning, thick blood film, microscopy automation, North East Nigeria.