Cargando…
Multiclass malaria parasite recognition based on transformer models and a generative adversarial network
Malaria is an extremely infectious disease and a main cause of death worldwide. Microscopic examination of thin slide serves as a common method for the diagnosis of malaria. Meanwhile, the transformer models have gained increasing popularity in many regions, such as computer vision and natural langu...
Autores principales: | , |
---|---|
Formato: | Online Artículo Texto |
Lenguaje: | English |
Publicado: |
Nature Publishing Group UK
2023
|
Materias: | |
Acceso en línea: | https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10564789/ https://www.ncbi.nlm.nih.gov/pubmed/37816938 http://dx.doi.org/10.1038/s41598-023-44297-y |
Sumario: | Malaria is an extremely infectious disease and a main cause of death worldwide. Microscopic examination of thin slide serves as a common method for the diagnosis of malaria. Meanwhile, the transformer models have gained increasing popularity in many regions, such as computer vision and natural language processing. Transformers also offer lots of advantages in classification task, such as Fine-grained Feature Extraction, Attention Mechanism etc. In this article, we propose to assist the medical professionals by developing an effective framework based on transformer models and a generative adversarial network for multi-class plasmodium classification and malaria diagnosis. The Generative Adversarial Network is employed to generate extended training samples from multiclass cell images, with the aim of enhancing the robustness of the resulting model. We aim to optimize plasmodium classification to achieve an exact balance of high accuracy and low resource consumption. A comprehensive comparison of the transformer models to the state-of-the-art methods proves their efficiency in the classification of malaria parasite through thin blood smear microscopic images. Based on our findings, the Swin Transformer model and MobileVit outperform the baseline architectures in terms of precision, recall, F1-score, specificity, and FPR on test set (the data was divided into train: validation: test splits). It is evident that the Swin Transformer achieves superior detection performance (up to 99.8% accuracy), while MobileViT demonstrates lower memory usage and shorter inference times. High accuracy empowers healthcare professionals to conduct precise diagnoses, while low memory usage and short inference times enable the deployment of predictive models on edge devices with limited computational and memory resources. |
---|