Redefining lameness assessment: Constructing lameness hierarchy using crowd-sourced data

Lameness causes pain to dairy cows and economic losses to farmers, but can be difficult to detect and routinely monitor. Despite numerous attempts to develop automatic detection methods, few have been successfully applied on farms. The development of reliable automated methods is likely restricted by the lack of large, labeled training datasets that capture the diversity in lameness cases within and among farms. Additionally, conventional gait scoring methods employed for annotating training videos are subjective and unreliable, adding noise to training data and thus hindering model performance. We propose a novel approach to lameness assessment in dairy cows, leveraging crowd-sourced data to construct a lameness hierarchy using the Elo-rating method. In this pilot study using 30 cow videos, our proposed lameness hierarchy constructed from pairwise lameness assessments achieved high inter-observer reliability (intraclass correlation coefficient (ICC) = 0.81) among experienced assessors. In contrast, we found that the traditional, subjective gait scoring systems to be inconsistent, with intra- and inter-observer reliabilities of ICC = 0.62±0.09 and 0.44±0.02, respectively. We also demonstrated feasibility for the pairwise assessment to be executed by untrained assessors (in this case, crowd workers recruited via Amazon MTurk), evidenced by high agreement between hierarchies generated by crowd workers and experienced assessors (ICC = 0.85). We created a subsampling algorithm, and found that recruiting just 8 crowd workers per video pair was sufficient to construct a reliable lameness hierarchy. This method also decreased the number of pairwise comparisons required by 61 %, relative to evaluating all possible comparisons between every pair of cows. We conclude that our proposed lameness hierarchy method, using easily accessible crowd workers to facilitate quick and accurate labeling of lameness videos, enables a reliable and granular evaluation of lameness. We suggest that this approach be used to create large training datasets suitable for developing robust lameness detection models.