Intelligent identification of medical and veterinary intracellular protozoa by using self-supervised learning

Zoonotic diseases pose a major threat to both human and animal health, contributing significantly to global morbidity and mortality. Accurate diagnosis is crucial for effective control and treatment, with microscopic examination serving as the gold standard, supplemented by highly sensitive molecular biology techniques. However, these confirmatory methods require skilled personnel and are subject to inter- and intra-rater variability. An innovative solution lies in artificial intelligence (AI)-powered automated tools, which offer a promising alternative. This study aimed to develop a self-supervised learning (SSL) approach using the Distillation with No Labels (DiNOv2) algorithm to extract features of protozoa from Giemsa-stained blood samples.

The development of self-supervised learning algorithms, including DiNOv2, was based on archived samples of clinically significant and significant veterinary microorganisms. These models were evaluated in comparison to a baseline Vision Transformer (ViT).

Among the tested SSL models, the DiNOv2-Small version achieved exceptional performance, surpassing 99% accuracy and specificity while maintaining the lowest misclassification rate (0.263). It also demonstrated a high area under the curve (AUC) value of 0.990, underscoring its robust classification capability. Remarkably, even when trained on only 20% of the dataset, the SSL models retained performance levels comparable to baseline models trained on the full dataset. However, further reducing the sample size below 20% led to notable declines in evaluation metrics: accuracy dropped by 5-7%, recall decreased from 37.9% to 35.1%, precision fell from 35% to 25.2%, and the F1 score declined from 31.1% to 21.5%. Additionally, the AUC decreased by 4-11%, while the misclassification rate increased, indicating reduced robustness. A key limitation of this study was the highly imbalanced fine-tuned dataset between classes. Nevertheless, the inconsistent performance observed may be mitigated by employing the larger DiNOv2 model, which improves the F1 score and enhances the model's ability to handle imbalanced data while reducing reliance on labeled data.

The proposed method can assist laboratory technicians, particularly in resource-limited healthcare settings. Furthermore, the findings support the potential deployment of this AI-based tool for automated screening in both medical and veterinary applications.