Udder thermogram-based deep learning approach for mastitis detection in Murrah buffaloes

Abstract

Mastitis, a production disease with multiple etiologies, inflicts significant economic losses among dairy farmers around the globe. In this study, an attempt has been made to detect mastitis through a Convolutional Neural Networks (CNN)-based deep learning model using 7615 udder thermograms of 40 Murrah buffaloes. The thermograms were grouped separately as healthy, sub-clinical (SCM), and clinical mastitis (CM) affected udder quarters based on California Mastitis Test (CMT) scores, Somatic Cell Count (SCC) values and thermal image analysis. Results of thermogram analysis revealed a significant increase (p < 0.01) in the mean values of udder skin surface temperature (USST) among SCM and CM-affected quarters compared to healthy quarters to the tune of 1.32 and 2.61 °C, respectively. The USST showed a strong positive correlation with the CMT score (r = 0.87, p < 0.01) and log10SCC value (r = 0.88, p < 0.01). The sequential (Normal vs. Clinical and Normal vs. Sub-clinical) models had training accuracy and validation accuracy of 0.999 and 0.988, 0.991 and 0.978, respectively. The confusion matrix for Normal vs. Clinical and Normal vs. Sub-clinical models reflected a loss of 0.009 and 0.029, precision of 0.947 and 0.980, and recall of 0.996 and 0.904, respectively. Consequently, the sequential (Normal vs. Clinical and Normal vs. Sub-clinical) models achieved a testing accuracy of 0.970 and 0.943, respectively. Thus, the improved deep-learning CNN models efficiently predicted SCM and CM cases in Murrah buffaloes.