A Deep Hybrid Model For Automated Malaria Parasite Identification And Stage Prediction

Abstract

Reliable detection as well as classification of malaria parasites in microscopic blood stain images are crucial for the planning of treatment and diagnosis. While classification-only systems can identify parasite species and stages from cropped regions, practical deployment requires detection-first models that localize parasites before classification. In this work, we evaluate three representative deep learning architectures—YOLOv8, DenseNet121, and CNN—on the MP-IDB dataset for combined parasite detection and stage classification. YOLOv8 is employed as a single-shot object detector for parasite localization, while DenseNet121 and CNN serve as classifiers on cropped regions of interest. A standardized training pipeline is adopted with preprocessing, augmentation, transfer learning, and stratified 5-fold cross-validation. Evaluation includes accuracy, F1-score, mean average precision (mAP), ROC-AUC, model size, and inference latency. Results show that YOLOv8 achieves strong localization performance with low latency, while DenseNet121 and CNN provide high classification accuracy on localized crops. Our study establishes a reproducible benchmark for hybrid detection-classification pipelines and provides practical guidance for deploying deep learning-based malaria diagnosis tools in low-resource environments.