PlasmoStage: A Hierarchical Deep Learning Framework for Plasmodium Parasite Staging in Malaria

Abstract

Malaria remains one of the most life-threatening infectious diseases worldwide. Recent advancements in deep learning and computer vision have shown significant promise in automating the analysis of medical images, including malariainfected blood smears. While hierarchical classification, a machine learning approach that organizes classes into a hierarchy from broad categories to specific subtypes, has proven effective in various domains, its application to malaria parasite staging at the single-cell level remains underexplored. In this work, we present PlasmoStage, a novel hierarchical deep learning framework for hierarchical malaria parasite staging. Our approach leverages the DinoBloom foundation model, a state-of-the-art selfsupervised model for single-cell image analysis in hematology, as a robust feature extractor. By fine-tuning these features using fully connected layers, PlasmoStage surpasses traditional flat classification models and baseline hierarchical methods. Our contributions are threefold: (1) We introduce a biologically inspired hierarchical classification framework that improves diagnostic accuracy and interpretability by aligning with the natural progression of malaria parasites. (2) We demonstrate the efficacy of foundation model-based feature extraction, achieving stateof-the-art performance with minimal fine-tuning. (3) We provide a comprehensive evaluation on publicly available datasets, including ablation studies and benchmarking against existing methods. Experimental results demonstrate that PlasmoStage effectively differentiates between uninfected and infected cells (Accuracy = 98.66%, F1-score = 98.38%), accurately identifies parasite species—Plasmodium falciparum or Plasmodium vivax (Accuracy = 98.75%, F1-score = 98.99%), and outperforms conventional flat and hierarchical classification approaches in parasite staging (Vivax staging: Accuracy = 85.90%, F1-score = 85.71%; Falciparum staging: Accuracy = 96.92%, F1-score = 96.62%). This work highlights the potential of leveraging foundation models for automated malaria diagnosis and staging.