Enhanced YOLO-based framework and benchmarking for automated Plasmodium vivax detection

Abstract

Despite remarkable progress in global health and increased life expectancy in the 21st century, malaria remains a major parasitic disease-causing significant morbidity and mortality, particularly in sub-Saharan Africa and Southeast Asia. Traditional diagnostic methods such as rapid diagnostic tests (RDTs) and microscopy, though widely used, suffer from limitations including biomarker mutations, reduced sensitivity at low parasite counts, and operator dependency. These challenges underscore the need for intelligent, automated diagnostic solutions. Artificial intelligence (AI)-driven computer vision methods, especially deep learning (DL)-enabled object detection architectures, offer promising alternatives for automated parasite identification. This study aims to perform a comparative evaluation of YOLO variants (v3, cascade v3, scaled v4, v5, and v8) to determine the most suitable architecture for precise and efficient detection of malaria. The models were assessed with following parameters such as, precision, accuracy, F1-score, recall and mean Average Precision (mAP). This study proposes a novel DL framework that combines YOLOv3 with a modified MobileNetV2 backbone, augmented by a Transformed Convolutional Layer (TCL)exhibits multi-scale texture sensitivity and efficient feature extraction, making it particularly effective for analyzing dense thick smear images of Plasmodium vivax (P. vivax).These findings provide guidance for deploying AI-based malaria diagnostic tools across varied clinical and field settings.This approach aims to contribute toward developing a robust, scalable, and interpretable diagnostic framework for early malaria detection and improved disease management.