Deep learning-based detection and viability assessment of Eimeria oocysts

TL;DR

This study uses deep learning to quickly and accurately assess the viability of Eimeria oocysts, which could improve coccidiosis vaccine development and management in poultry.

Contribution

A deep learning model based on YOLOv7 was developed to distinguish viable from non-viable Eimeria oocysts using morphological features.

Findings

The model trained with phase-contrast images achieved 93.1% precision and 91.2% recall for viability detection.

Refined labeling improved model performance to 99.1% precision and recall.

The model generalized well to E. tenella but required fine-tuning for E. maxima due to size differences.

Abstract

Coccidiosis, caused by Eimeria species, is a significant disease affecting the poultry industry worldwide, leading to substantial economic losses due to reduced flock performance. Effective vaccination strategies require the precise quantification of the dosage of viable Eimeria oocysts to induce immunity in young chicks without causing disease. However, current methods for determining oocyst viability rely on sophisticated equipment and are not effective for routine monitoring. Recently, we documented the presence of granular structures exclusively in dead oocysts using high-resolution microscopic imaging. Hence, this study aimed to develop a simple, cost-effective approach using deep learning-based models to distinguish viable from non-viable Eimeria oocysts using morphological features, including the presence/absence of granular structures. Phase-contrast (PC), differential interference contrast (DIC), and brightfield (BF) imaging were employed to capture E. acervulina oocysts. The performance of a deep convolutional neural network based on the YOLOv7 architecture was evaluated for viability detection. Results indicated that the model trained with PC images outperformed those trained with DIC and BF, achieving overall precision and recall of 93.1 % and 91.2 %, respectively. Further dataset refinement, including class-specific labeling for sporulated, unsporulated, and dead oocysts, enhanced model performance, achieving an overall precision and recall of 99.1 % and 99.1 %, respectively. Cross-species evaluation of the method demonstrated that the model trained on E. acervulina generalized well to E. tenella, achieving 100 % overall precision and 98.1 % recall without additional training, whereas initial cross-species performance for E. maxima was substantially lower (43.5 % of overall recall), likely due to its larger oocyst size, but exceeded 95 % accuracy after fine-tuning with an E. maxima-specific dataset. This study highlights the potential of deep learning approaches to provide a practical, rapid, and reliable method for evaluating Eimeria oocyst viability, contributing to improved vaccine formulation and better coccidiosis management in the poultry industry. This proof of principle may also find application in assessing the viability of related parasites, such as Cyclospora cayetanensis, that pose a risk to human health and food safety.

Figures