Poultry diseases diagnostics models using deep learning.

Area of Science:

• Veterinary Medicine
• Artificial Intelligence
• Agricultural Science

Background:

• Poultry diseases such as Coccidiosis, Salmonella, and Newcastle significantly impact poultry production.
• Limited access to agricultural support services in regions like Tanzania hinders early disease detection.
• Deep learning offers a promising approach for the early diagnosis of poultry diseases.

Purpose of the Study:

• To develop and evaluate a deep Convolutional Neural Network (CNN) model for the early diagnosis of poultry diseases using fecal image classification.
• To compare the performance of various CNN architectures (VGG16, InceptionV3, MobileNetV2, Xception) in diagnosing poultry diseases from fecal images.

Main Methods:

• Collected and labeled 1,255 laboratory fecal images and 6,812 farm-labeled fecal images using Open Data Kit.
• Trained and fine-tuned baseline CNN, VGG16, InceptionV3, MobileNetV2, and Xception models on the collected fecal image datasets.
• Evaluated model performance using farm-labeled images as the test set, with a focus on accuracy and F1 scores.

Main Results:

• Models achieved high test accuracies without fine-tuning, with InceptionV3 reaching 94.79%.
• Fine-tuning, particularly with a frozen batch normalization layer, significantly improved accuracies, with Xception achieving 98.24% and MobileNetV2 reaching 98.02%.
• All fine-tuned models demonstrated F1 scores above 75% across all four classes (healthy, Coccidiosis, Salmonella, Newcastle).

Conclusions:

• Deep learning models, especially fine-tuned CNNs, are highly effective for diagnosing poultry diseases from fecal images.
• MobileNetV2 is recommended for farm-level deployment due to its lighter weight and superior generalization ability.
• This approach can significantly aid in the early detection of poultry diseases, thereby improving poultry production and farmer livelihoods.