Detection of Newcastle Disease in Chicken using Xception-XGBoost Framework

Abstract

Newcastle disease or Newcastle (ND) is a serious viral disease of poultry. The high contagious rate and economic cost is mainly alarming due to its high mortality rate and negative economic impact on farmers. This study put forward a hybrid framework based on the poultry (chicken) faecal imagery to detect Newcastle disease. The framework (Xception+XGBoost) extracts features using a pre-trained Xception CNN and classifies the diseased samples using an XGBoost classifier. For tackling the class imbalance issue between ND and Non-ND (Healthy and other diseased samples) multiple image augmentation techniques, class-weighted loss function, and feature-level oversampling of ND samples were used. It was specifically tuned to classify accurate diseases diagnosis with a decision threshold of 0.429. The cross-validation results show an overall accuracy of 98.10% set in a held test set. The model showed a recall of 88.24% (75 true positive cases were detected out of 85) and precision of 85.23% for the minority Newcastle disease class and an F1-score of 86.71%. The further evaluation resulted a promising Precision-Recall Area Under the Curve (PR-AUC) of 95.50%, which indicates potential to tackle class imbalance. The results suggest that the suggested structure offers effective solutions to the poultry disease detection and can be a useful computer-aided screening tool.