A Two-Stage Deep Learning Approach with Image Tiling for Malaria Diagnosis from Microscopic Blood

Abstract

Malaria diagnosis is a severe difficulty in preserving human health, particularly in developing nations. It is essential if correct treatment and control measures are to be instituted and implemented. Thus, this research focuses on A Two-Stage Deep Learning Approach with Image-Tiling for Malaria Diagnosis from Microscopic Blood Smears. Due to the necessary image downscaling, traditional deep-learning models for malaria screening in high-resolution microscopic images often lose crucial spatial details. This research introduces an image-tiling technique with overlaps to preserve spatial resolution and enhance detection accuracy. Comparing YOLOv9 and RT-DETR, the study aims to identify which models are the most suitable for accurate malaria detection in blood smears, focusing on eliminating false positives and enhancing the model's reliability. This method represents a significant development since it preserves the integrity of high-resolution data throughout the detection procedure. The researchers propose a thorough two-stage detection methodology employing sophisticated tiling and refined object identification algorithms to improve malaria diagnosis. Images are divided into 640x640 tiles, and overlapping between tiles is 30% so as not to lose spatial information. The first stage utilizes the YOLOv9 algorithm to identify malaria parasites with an accuracy of 0.896. The MobileViT model subsequently enhances these detections by minimizing false positives, attaining a validation accuracy of 0.9833. After establishing malaria's existence, the procedure advances to the second stage, where thin blood smears are examined to identify kinds of malaria parasites using identical models, achieving detection accuracies of 0.931 and classification accuracies of 0.929188. This two-part approach significantly improves diagnostic accuracy and decreases the time needed to arrive at a diagnosis. The culmination of this method is integrated into a fully autonomous desktop application designed to seamlessly mesh with existing healthcare systems, dramatically reducing diagnostic response times. This novel application enhances patient quality by providing quick and accurate malaria diagnosis. It enhances the possibilities of deploying deep learning models in medical imaging, which promises to transform patient outcomes in several medical fields.