Research Article: Deep learning–assisted malaria microscopy with sensitivity-aware threshold optimization
Abstract:
Malaria microscopy becomes clinically risky when parasitized erythrocytes are missed, even when a classifier reports high overall accuracy. This study presents an artificial-intelligence-assisted malaria microscopy framework that explicitly optimizes the diagnostic operating point for sensitivity-aware screening. A balanced set of 27,558 National Institutes of Health/National Library of Medicine (NIH/NLM) thin-smear cell images was divided by stratified image-level sampling in the ratio 70:15:15 for training, validation, and independent testing. Two custom convolutional neural networks (CNNs) with complementary capacity–regularization profiles and an equal-weight score-level ensemble were evaluated. Training-only online augmentation, fixed input resizing, validation-only threshold selection, confidence intervals, and error analysis were incorporated to improve reproducibility. At the conventional threshold, the compact CNN achieved the highest accuracy (95.26%). After optimizing the operating point with the recall-weighted F 2 objective, the deeper CNN provided the strongest screening-oriented result on the held-out test set, reaching 97.05% sensitivity (95% confidence interval: 96.23%–97.70%), a 96.01% F 2 -score, a 2.95% false-negative rate, and a receiver-operating-characteristic area under the curve of 0.9876 (95% confidence interval: 0.9842–0.9910). It reduced missed parasitized cells from 177 to 61 relative to its default threshold. The public release does not provide patient identifiers; therefore, the reported split is image-level rather than confirmed to be patient-level independent, and external multicenter validation remains necessary before clinical deployment. The main finding is that the highest-accuracy classifier is not necessarily the most appropriate sensitivity-oriented screening configuration.
Introduction:
Malaria remains a major global health burden. The World Health Organization estimated 282 million cases and 610,000 deaths in 2024, underscoring the continuing need for timely diagnosis and sustained surveillance ( 1 ). Light microscopy remains central to malaria diagnosis, but its performance can be influenced by parasite density, staining quality, field selection, workload, and reader experience. Automated image analysis is therefore most appropriately considered as a decision-support tool that can assist,…
Read more