Onderwerp scriptie

The effect of different observers as golden standard on the performance of an automatic lameness detection system – Lisette van Zuijlen

Automatic lameness detection systems (ALDS) have been developed in order to help farmers in lameness control of their cattle. One of the requirements for the development of such ALDS is a valid and reliable reference or golden standard (GS) able to detect lameness. Since most of the ALDS use one observer as GS, the aim of this study was to investigate how different observers used as GS affect the output and performance of an ALDS.

Farmers are aware of 25 to 50 % of the lame cows and early signs of lameness remain most of the time undetected. In order to help in lameness control, automatic lameness detection systems (ALDS) have been developed. One of the requirements for the development of such ALDS is a valid and truly reliable reference or golden standard (GS) able to detect lameness. Since most of the ALDS use one observer as GS, the aim of this study was to investigate how different observers used as GS affect the output and performance of an ALDS based on the detailed posture trait arched back. 

For this study data were used coming from an experiment performed on a dairy farm located in Yifat (Israel). Cows were simultaneously recorded by a 2D and a 3D camera while they walked through an alley when leaving the milking area. The 2D videos (58) were scored twice in two different sessions by 10 observers on 5 individual gait/posture traits and the overall locomotion. The scores provided by each of the 10 observers were considered as individual GS for the calibration of the ALDS. The 3D videos were used to run the ALDS. The performance of the ALDS was determined by the sensitivity, specificity and accuracy.  

In accordance with literature, locomotion scores and gait/posture trait outcomes varied between observers. Since different observers were used as GS the expectation was that the ALDS output would be similar as the observer output. This, however, was not the case. This could be explained by the different backgrounds, skills and interpretations of the observers regarding locomotion scoring. The sensitivity ranged from 0.71 to 1, the specificity ranged from 0.75 to 1 and the accuracy ranged from 0.81 to 0.96. The sensitivity was higher than the specificity indicating that the ALDS was slightly better detecting lame cows instead of not lame cows. Different observers as GS did have an effect on the performance and the output of the ALDS.

Student: EAM van Zuijlen

Supervisor: dr ir E Bokkers (APS), A Schlageter Tello, MSc (ASG)

24 Ects