Hierarchical Multi-Head Deep Learning for Robust Malaria Detection in Blood Smear Images

Abstract

Microscopy-based malaria diagnosis remains the primary standard in many endemic regions; however, its accuracy depends heavily on blood smear quality and the expertise of microscopists. Most current deep learning approaches perform direct cell classification without incorporating clinical examination workflows, which increases vulnerability to errors caused by poor smear quality and morphological similarities among blood cell types. This study presents a Hierarchical Multi-Head Deep Learning framework that replicates the clinical microscopy workflow through sequential blood cell type identification, erythrocyte quality assessment, and gated parasite infection detection. The model employs a shared backbone with multiple prediction heads, ensuring that infection detection is limited to relevant cells via a hierarchical gating mechanism. Experiments using combined malaria cell datasets and general blood cell datasets, including simulated smear degradation scenarios, demonstrate that the model achieves F1-scores of 0.987 for cell type classification, 0.944 for quality assessment, and 0.958 for malaria infection detection, while significantly reducing false positives on white blood cells compared to single-task classifiers. These results indicate that integrating clinical diagnostic workflows into deep learning architectures improves the robustness and reliability of automated malaria detection systems, especially for deployment in resource-limited settings. Furthermore, the proposed framework enhances interpretability by aligning model decisions with clinically meaningful stages, enabling better integration into real-world diagnostic pipelines. This alignment also supports improved generalization under variable imaging conditions commonly encountered in field microscopy.