A deep architecture based on attention mechanisms for effective end-to-end detection of early and mature malaria parasites

• Published in: Biomedical signal processing and control Vol. 94; p. 106289
• Main Authors: Zedda, Luca, Loddo, Andrea, Di Ruberto, Cecilia
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01-08-2024
• Subjects: Computer vision; Deep learning; Early malaria diagnosis; Image processing; Malaria parasites detection; Deep learning; Computer vision; Image processing; Early malaria diagnosis; Malaria parasites detection
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2024.106289

Malaria is a severe infectious disease caused by the Plasmodium parasite. The early and accurate detection of this disease is crucial to reducing the number of deaths it causes. However, the current method of detecting malaria parasites involves manual examination of blood smears, which is a time-consuming and labor-intensive process, mainly performed by skilled hematologists, especially in underdeveloped countries. To address this problem, we have developed two deep learning-based systems, YOLO-SPAM and YOLO-SPAM++, which can detect the parasites responsible for malaria at an early stage. Our evaluation of these systems using two public datasets of malaria parasite images, MP-IDB and IML, shows that they outperform the current state-of-the-art, with more than 11M fewer parameters than the baseline YOLOv5m6. YOLO-SPAM++ demonstrated a substantial 10% improvement over YOLO-SPAM and up to 20% against the best-performing baseline in preliminary experiments conducted on the Plasmodium Falciparum species of MP-IDB. On the other hand, YOLO-SPAM showed slightly better results than YOLO-SPAM++ in subsets without tiny parasites, while YOLO-SPAM++ performed better in subsets with tiny parasites, with precision values up to 94%. Further cross-species generalization validations, conducted by merging training sets of various species within MP-IDB, showed that YOLO-SPAM++ consistently outperformed YOLOv5 and YOLO-SPAM across all species, emphasizing its superior performance and precision in detecting tiny parasites. These architectures can be integrated into computer-aided diagnosis systems to create more reliable and robust systems for the early detection of malaria.