Detecção de Ovos de Parasito em Imagens de Microscópio via Ensemble de Detectores

Abstract

Intestinal parasitic infections represent a global health problem, especially in regions with precarious sanitary conditions and tropical areas. According to the World Health Organization (WHO), it is estimated that 1.5 billion people are affected by these infections. The diagnostic manual, carried out by professionals who analyze microscopic samples to identify and classify parasitic eggs, is time-consuming and susceptible to errors, mainly due to variation in professionals’ experience and the presence of small parasitic structures or those partially hidden by impurities. To address this challenge, this work proposes an ensemble-based methodology that combines a specialized detection architecture with generic detectors refined on a specific image set. The custom architecture, developed from HRNet-W32, incorporates residual blocks and loss composition. Detectors such as Cascade R-CNN, TOOD and CenterNet were fine-tuned using the Chula-ParasitEgg11 dataset. The fusion of detections was carried out using the Weighted Box Fusion technique, which integrates results without discarding possible relevant identifications. The approach achieved significant performance, with an F1-Score of 0.945 and mAP[0.5:0.95] of 0.868, demonstrating the potential to automate diagnosis accurately and efficiently.