Constructing a Predictive Model for STH and Schistosomiasis Classification From Microscopic Images

Soil‐transmitted helminths (STHs) and schistosomiasis are widespread parasitic diseases in tropical regions, particularly in Africa, with substantial health and socioeconomic burdens. Early diagnosis and treatment are critical for mitigating these impacts. Conventional microscopy‐based diagnosis was time‐consuming and labor‐intensive, posing challenges in resource‐limited settings such as Ethiopia. This study developed an innovative system that combined machine learning (ML) and deep learning to analyze microscope images of parasite eggs, improving diagnostic speed and accuracy compared to traditional CNN‐only approaches. We compared a hybrid CNN–ML approach with standalone deep learning models and vision transformers (ViTs) for classifying five categories: Ascaris , hookworm, schistosomiasis, Trichuris , and negative samples. The dataset comprised 1490 images from the Ethiopian Public Health Institute, processed with resizing, normalization, and augmentation. CNN architectures (VGG16, ResNet50, DenseNet121, MobileNetV2, and EfficientNetB0) and ViT served as feature extractors, with ML classifiers (SVM, XGBoost, KNN, RF, and DT) performing the predictions. The hybrid CNN–ML model outperformed standalone models, with VGG16‐SVM and VGG16‐XGBoost achieving the highest test accuracy of 99.31% and 99.35%, respectively. In contrast, standalone CNNs showed lower accuracy (VGG16: 79.98%; DenseNet121: 84.12%). Negative samples were classified with high accuracy across models, while parasite classes exhibited varying performance depending on the architecture. This system enhances diagnostic utility in low‐resource settings by enabling real‐time analysis. However, limitations include a small, long‐stored dataset with limited diversity and potential degradation, which may affect model generalizability.