Malaria remains a major global public health challenge, transmitted through the bites of infected female mosquitoes. Despite progress in prevention and treatment, the disease continues to cause significant morbidity and mortality, especially among children and pregnant women. Conventional diagnostic methods relying on microscopic examination are time-consuming, dependent on skilled personnel, and prone to human error. Existing deep learning models for malaria detection often show limited performance and overfitting. A hybrid deep learning model was developed by combining EfficientNetB0 with a custom convolutional neural network (CNN) enhanced by attention mechanisms to classify malaria-infected and uninfected red blood cell images. EfficientNetB0 provided pre-trained global feature extraction, while the custom CNN captured domain-specific features. The model was trained and evaluated using a publicly available malaria dataset. The proposed hybrid model achieved a classification accuracy of 96.53%, precision of 94.80%, recall of 98.22%, F1-score of 95.95%, and an AUC of 99.12%. Attention maps offered interpretability by highlighting biologically relevant regions within cell images. Comparative experiments showed that the hybrid model outperformed standalone EfficientNetB0 and CNN architectures. Integrating global and domain-specific feature representations significantly improves malaria image classification performance. The proposed hybrid model demonstrates strong potential for use in automated malaria diagnosis systems, supporting early detection and timely treatment. • The hybrid deep learning model achieved exceptional results with 96.53% accuracy, 98.22% recall, and 99.12% AUC. • The research demonstrates that a two-phase fine-tuning approach dramatically outperforms single-phase training. • The study introduces an innovative combination of EfficientNetB0 with a custom CNN. • Integration of attention mechanisms and Grad-CAM visualisation allows the model to highlight biologically relevant regions in cell images.
Oladimeji et al. (Sun,) studied this question.