Desenvolupament d'un sistema de diagnòstic automatitzat per la malària i l'esquistosomiasi urogenital utilitzant eines d'intel·ligència artificial i un microscopi robotitzat universal de baix cost

Carlos Rubio Maturana

Dr. Joan Joseph Munné (VHIR) and Dra. Elisa Sayrol Clols (UPC)

Malaria caused 263 million infections worldwide in 2023 according to the World Health Organization (WHO). Schistosomiasis is classified by the WHO as a neglected tropical disease, with more than 253 million people at risk of infection. Microscopy remains the gold standard technique for the diagnosis of both diseases, however it is a professional-dependent method, which requires long observation times under the microscope. As an alternative to traditional methods, new diagnostic techniques based on artificial intelligence are being developed, which allow for rapid, accurate and automated diagnosis. In the presented project, a new diagnostic system based on a smartphone application and a low-cost robotic microscope have been developed. The system is capable of performing a fully automated diagnosis by: (i) scanning the biological sample and autofocusing the microscopic fields for digital image acquisition with the smartphone, (ii) analysis of the digital images using neural network models trained for parasite and cell detection, and (iii) final technical/optional validation of the diagnosis using LIS software. The main objective of the study is the development and validation of the system, especially designed for its implementation in clinical laboratories, including resource-poor settings where these diseases are endemic. 148 Giemsa-stained thick-drop samples were used, of which 2571 digital images were captured and labeled. The comparative analysis of neural networks demonstrated an optimal performance of the YOLOv5x model, with an F-score of 92.79% for the detection of immature, mature Plasmodium trophozoites and leukocytes. For urogenital schistosomiasis, a total of 1017 labeled digital images of 24 urine sediment samples were analyzed. The training of the YOLOv5x model demonstrated an F-score of 99.30% for the detection of S. haematobium eggs, and the NasNetLarge binary neural network an accuracy of 85.60% for the detection of erythrocytes/leukocytes in urine. The diagnostic algorithms were integrated into a smartphone application and a LIS laboratory software. With the results obtained, the system is able to determine whether a thick blood drop sample is positive/negative for Plasmodium, and its parasitemia levels; and to detect the presence of S. haematobium eggs and erythrocytes/leukocytes in urine sediment samples. A diagnostic validation of the system was performed to compare it with the gold standard technique of conventional microscopy. The results obtained at the Drassanes-Vall d'Hebron Center (Barcelona, ​​Spain) demonstrated a sensitivity of 81.25% and a specificity of 92.11% for the diagnosis of malaria. Additionally, a proof of concept was performed at the Nossa Senhora da Paz Hospital (Cubal, Angola), obtaining satisfactory qualitative results for the implementation of the system in laboratories with few resources. The project has a holistic approach with a transversal impact on global health. The coalescence of the automation of the diagnostic process and the accessibility of the system give the prototype the optimal characteristics to join the global effort to fight malaria and neglected tropical diseases.

The thesis by Dr. Carles Rubio Maturana focuses on developing an automated diagnostic system for malaria and urogenital schistosomiasis using artificial intelligence (AI) and a low-cost universal robotized microscope. Key aspects include: Diagnostic Innovation: Combines digital image analysis with convolutional neural networks (CNNs), primarily YOLO models, for real-time detection of parasites in blood and urine samples. Hardware Development: Open-source 3D-printed components and servo motors adaptable to conventional optical microscopes, controlled via Arduino and a mobile app. Performance: Malaria detection: YOLOv5 achieved ~92% precision and recall. Schistosomiasis detection: YOLOv5 achieved ~99% precision and recall. Validation: Conducted in hospital labs and in endemic regions (Angola), assessing accuracy, usability, and implementation feasibility in low-resource settings. Accessibility: System is low-cost, open-source, and does not require internet connectivity, making it suitable for remote areas. Additional Applications: Training microscopists. Quality control of staining techniques. Laboratory information management via mobile devices. Impact: Improves diagnosis of neglected diseases. Strengthens public health systems. Promotes open science through shared datasets and code. Commercial and Social Potential: Addresses global health challenges with scalable, affordable technology.

Malaria, Schistosomiasis, Urogenital schistosomiasis, Automated diagnosis, Artificial intelligence, AI, Digital microscopy, Image analysis, Convolutional neural networks, CNN, YOLO, YOLOv5, Object detection, NasNetLarge, Thick blood smear, Giemsa stain, Urine sediment, Parasite detection, Plasmodium spp, Schistosoma haematobium, Mobile application, Arduino, Servo motors, 3D printing, Open-source hardware, Robotized microscope, Low-cost technology, Remote diagnostics, Endemic regions, Laboratory automation, Real-time analysis, Diagnostic accuracy, Sensitivity, Specificity, F-score, mAP, SafeCreative registration, Creative Commons license, Open science, Data sharing, CORA-CSUC repository, Training microscopists, Quality control, Staining anomalies, Laboratory information systems, LIS, Mobile data management, Public health, Neglected diseases, Technological transfer, Global health, Accessibility, Implementation feasibility.