Interpretable deep ensemble model for species-level malaria detection from thin blood smears
Firehiowt Belay, Hussien Seid, Abeba Gebretsadik Reda, Bokretsion G. Brhane, Geremew Tasew · BMC Medical Imaging · 2026
Malaria remains a major global health concern, particularly in sub-Saharan Africa, where accurate species-level identification is essential for guiding targeted treatment and elimination strategies. This study introduces a deep learning ensemble model that employs a soft-voting mechanism to combine a custom convolutional neural network (CNN) with five established architectures for reliable malaria parasite detection and species differentiation from thin blood images. The study utilized under real-world diagnostic conditions in Ethiopia, a soft-voting ensemble framework combining a custom convolutional neural network (CNN) with VGG16, DenseNet121, InceptionV3, MobileNetV2, and Xception. The model was developed, trained, and evaluated using 37,200 Giemsa-stained thin blood smear images collected in Ethiopia, reflecting real-world diagnostic conditions. The ensemble consistently outperformed individual models, achieving classification accuracies ranging from 99.87% to 99.987%, with a macro-averaged F1-score of 0.9822 (± 0.0028). Class-wise AUC-ROC values exceeded 0.9991, and the Brier score of 0.016 indicated well-calibrated probability estimates. Additional performance metrics, including Cohen’s Kappa (0.9856) and Matthews Correlation Coefficient (MCC) (0.9857), demonstrated strong inter-rater agreement. Interpretability was assessed using Grad-CAM visualizations, which revealed high spatial alignment with diagnostic features. Intersection-over-union (IoU) scores were 0.85 for P. falciparum, 0.82 for P. vivax, and 0.91 for negative samples. Notably, 91% of accurate predictions with clustering thresholds above 0.7 exhibited biologically meaningful IoU values. Misclassifications and false positives in negative smears showed limited spatial focus (IoU = 0.14). The proposed framework offers a scalable, interpretable, and clinically relevant solution for malaria diagnosis in endemic regions. Suggesting morphological plausibility. While the model demonstrates strong potential for field deployment, limitations include reliance on thin smears and lack of validation for mixed-species infections. Future work will address IoU-aware calibration and mobile deployment strategies.