Post-mortem detection of unhealthy livers and hearts in chickens using deep learning, logistic regression and Computed Tomography (CT) scanning

Abstract

Diseased chicken organs, including livers and hearts, are frequently observed during post-mortem inspection in poultry slaughterhouses. It is crucial to accurately identify and discard these organs, since they are unfit for human consumption. The current method of inspection is based on visual human examination and this procedure has limitations that negatively affect its reliability e.g. subjectivism and working under time pressure. It implies that new technologies should be investigated. Computed Tomography (CT) scanning can detect pathologically changed organ tissues, because the radiodensity of the organs is altered. Therefore, the aims of this study were to compare the radiodensity between healthy and diseased hearts and livers and then develop different classifiers to identify diseased organs. 264 chicken hearts and 252 livers were collected from a slaughterhouse including healthy and diseased samples. All the organs were CT scanned in a veterinary clinic. Two logistic regression models (Log_Reg_Hearts and Log_Reg_Livers) and four deep learning models were developed including deep and shallow neural networks (NN_Deep_Hearts/Livers and NN_Shallow_Hearts/Livers) to classify these organs. Deep learning models for hearts (accuracy 0.91 for NN_Shallow_Hearts; 0.92 for NN_Deep_Hearts) outperformed logistic regression model (0.82, Log_Reg_Hearts). There was no difference in accuracy for the liver models (0.78 for NN_Shallow_Livers; 0.75 for NN_Deep_Livers; 0.79 for Log_Reg_Livers). Our study confirms that diseased chicken hearts and livers can be automatically and accurately detected using CT scans classified by logistic regression/deep learning models. Overall, CT scanning has potential to increase the safety of poultry edible organs and streamline the workflow in the slaughterhouse.