Feasibility of the use of deep learning classification of teat-end condition in Holstein cattle

Abstract

Infections with pathogenic bacteria entering the mammary gland through the teat canal are the most common cause of mastitis in dairy cows; therefore, sustaining the integrity of the teat canal and its adjacent tissues is critical to resist infection. The ability to monitor teat tissue condition is a key prerequisite for udder health management in dairy cows. However, to date, routine assessment of teat condition is limited to cow-side visual inspection, making the evaluation a time-consuming and expensive process. Here, we demonstrate a digital teat-end condition assessment by way of deep learning. A total of 398 digital images from dairy cows' udders were collected on 2 commercial farms using a digital camera. The degree of teat-end hyperkeratosis was scored using a 4-point scale. A deep learning network from a transfer learning approach (GoogLeNet; Google Inc., Mountain View, CA) was developed to predict the teat-end condition from the digital images. Teat-end images were split into training (70%) and validation (15%) data sets to develop the network, and then evaluated on the remaining test (15%) data set. The areas under the receiver operator characteristic curves on the test data set for classification scores of normal, smooth, rough, and very rough were 0.778 (0.716-0.833), 0.542 (0.459-0.608), 0.863 (0.788-0.906), and 0.920 (0.803-0.986), respectively. We found that image-based teat-end scoring by way of deep learning is possible and, coupled with improvements in image acquisition and processing, this method can be used to assess teat-end condition in a systematic and efficient manner.