Identification of veterinary and medically important blood parasites using contrastive loss-based self-supervised...

Area of Science:

• Veterinary Parasitology
• Medical Diagnostics
• Machine Learning in Healthcare

Background:

• Zoonotic blood parasites pose significant global health risks to animals and humans.
• Traditional microscopic diagnosis is labor-intensive, time-consuming, and requires expert interpretation.
• There is a need for innovative, efficient, and accurate methods for parasite identification.

Purpose of the Study:

• To develop a self-supervised learning (SSL) approach for identifying zoonotic blood parasites from microscopic images.
• To focus initially on the classification of different parasite species.
• To enhance current diagnostic capabilities for blood-borne zoonotic diseases.

Main Methods:

• Utilized a public dataset of microscopic images of Giemsa-stained blood films.
• Employed the Bootstrap Your Own Latent (BYOL) algorithm for SSL model training.
• Tested Residual Network (ResNet) architectures (ResNet50, ResNet101, ResNet152) as backbones.
• Compared the performance of the SSL model against baseline supervised learning models.

Main Results:

• The BYOL SSL model demonstrated superior performance compared to supervised learning models across all parasite classes.
• ResNet50 backbone achieved high accuracy (0.992) in parasite classification.
• Fine-tuned SSL models reached 95% accuracy and 0.960 ROC AUC with only 1% of labeled data.
• SSL models trained with 20% of data achieved ≥95% across multiple metrics (accuracy, recall, precision, F1 score).

Conclusions:

• The developed SSL approach significantly enhances the accuracy and efficiency of zoonotic blood parasite identification.
• This method can improve disease surveillance, control, and outbreak prevention, especially in resource-limited settings.
• SSL effectively addresses challenges of data variability and extensive labeling requirements in biological and medical image analysis.