Deep learning and image processing for high-throughput udder phenotyping in dairy cows

Udder conformation is a crucial phenotype in dairy cows, as it relates to their health and productivity, particularly in automated milking systems (AMS). As such, this project aims to (1) develop an automated pipeline that extracts udder phenotypes from depth images and (2) evaluate the association between those traits and production performance. We recorded depth videos from the udders of 150 lactating Holstein cows at a farm with AMS and generated reports containing the production records of the cows' corresponding lactation. We trained, validated, and tested convolutional neural networks (CNN) for classification, segmentation, and keypoint detection using images from a subset of the AMS cows (n = 50) and an existing dataset (n = 63). The pipeline integrates CNN with watershed segmentation and image processing tools to select frames that portray the udder, segment the udder region (intersection over union = 0.93 ± 0.035), locate the teats (normalized Euclidean distance = 0.061 ± 0.06), divide the udder into its quarters, perform quality control, and extract phenotypes. These phenotypes include udder and quarter circularity, eccentricity, surface area, and ventral volume, as well as teat length and distance between teats. Using the pipeline, we extracted udder phenotypes from all AMS cows with good images and predictions (n = 138). We evaluated the association between the udder phenotypes and production performance using Pearson correlations and linear mixed models. Udders and quarters with larger area were associated with higher milk yields (r = 0.55). Similarly, cows with increased parity were associated with larger udder perimeter (r = 0.72), longer teats (r = 0.31), and larger distance between teats (r = 0.48). This pipeline is suitable for high-throughput udder phenotyping, which is relevant for breeding, health, management, and welfare.