Towards effective and efficient machine learning models for schistosomiasis diagnosis in microscopic images

Schistosomiasis is an infectious disease caused by the parasite Schistosoma mansoni, and it is a significant health concern in many underdeveloped tropical regions. The traditional method of diagnosing schistosomiasis involves fecal examinations under a microscope for parasite eggs, a process that is time consuming, error-prone, and requires specialized training. Artificial intelligence encompasses a broad range of techniques, including both machine learning and its subset, deep learning, which have been successful in solving such problems. Therefore, this study aimed to propose an automated solution that combines DL-based object detection methods with classical ML techniques and HOG feature extraction to identify S. mansoni eggs in images obtained using the Kato-Katz parasitological technique. A real database of 1100 images was created, processed, and annotated by three parasitologists. The proposed system achieved an Average Precision of 0.884 for an @[IoU = 0.50] using a Faster R-CNN method with ResNet-50 architecture over the best-annotated data, outperforming other detection models. Based on a threshold analysis, we suggest an integrated approach with a voting scheme of machine learning models to improve results in terms of false positives and false negatives. Comparative analysis with external data and a commercial tool confirmed that our approach offers promising results and applicability to assist health professionals in diagnosing schistosomiasis in public health systems, such as the Brazilian Unified Health System (SUS), potentially improving diagnosis speed and accuracy in endemic regions.