Research Project

Thomas Cummings - 3rd May 2024

This interdisciplinary research project combines the fields of computer science and data science as well as aspects from animal science with a focus on developing an automated machine learning system to classify canine body condition scores. This will be useful as a study by German et al. (2018) found that 74% of adult dogs were overweight or obese which is concerning given that overweight dogs “have a shortened life span, their quality of life is adversely affected, and they are predisposed to other conditions including osteoarthritis, diabetes mellitus and certain types of neoplasia”. It is interesting to consider that “90% of owners were aware of the health implications, but 93% were not aware of the existence of BCS charts or how to use them” (Eastland-Jones et al, 2014) and as such this shows a gap for an automated easy to use system.

“An automated body condition scoring system would be preferred to observational scoring because it would require less time, be less stressful on the animal, be more objective and consistent, and possibly be more cost effective” (Bewley et al, 2008). As a result this would have potential uses in monitoring stray dog populations where manual palpation wouldn’t realistically be feasible due to stress on the animals and causing a bite risk as well as in rehoming centres for new intakes prior to a full assessment. In addition to this dog owners could also make use of a system to bring more awareness to canine obesity and assist with monitoring dogs body classification between vet appointments. Having said this, it is important to note that the system should only be used as a guide and that a veterinarian should be consulted before making any changes to a dogs diet or exercise.

The primary aim of the research will be to practically demonstrate and support the hypothesis that a machine learning system can achieve strong levels of accuracy in determining body condition scores in the Labrador Retriever breed. Consideration will also be given in determining the theoretical generalisability of a machine learning approach across different breeds with the hypothesis that the Labrador Retriever model will have the greatest accuracy in morphologically similar breeds (figure 1.)

The model development phase of the project will first consider a morphological approach focusing on extracting measurements from the image in line with current research by Gant et al. (2016). This research suggested that abdominal to thoracic ratios provide a strong correlation to body condition score however other measured and calculated features will also be considered using statistical analysis. A second method will consider a deep learning approach using convolutional neural networks trained on labelled photos to identify features which determine the body condition score. Finally an ensemble method with a weighted application taken from both of the previously defined models will also be considered.

In order to develop these models, a significant amount of data for training, validation and testing datasets will be required primarily consisting of participant submitted photos both in the lateral and dorsal plane. These will then be manually labelled with the body condition score determined using a visual inspection approach. Further information regarding gathering and analysing the data requirements can be found here.

Once the models are trained and tested on the Labrador Retriever image set, statistical analysis will be used to verify the extent to which the models accurately classify body condition scores in order to support or reject the initial hypothesis. Further testing can then be completed on the various other breed images to determine the model generalisability. Further information regarding measuring the success of the project can be found here.

In conclusion the research project will aim to produce a functioning machine learning system for Labrador Retrievers and in doing so will determine that a system is feasible. As such this will provide the foundation for further research allowing for development of a generic non breed specific system with the aim of having a positive impact on general canine health.

References

Bewley, J.M., Peacock, A.M., Lewis, O., Boyce, R.E., Roberts, D.J., Coffey, M.P., Kenyon, S.J., & Schutz, M.M. (2008). ‘Potential for estimation of body condition scores in dairy cattle from digital images’, Journal of Dairy Science, 91(9), pp. 3439–3453. https://doi.org/10.3168/jds.2007-0836

Gant, P., Holden, S.L., Biourge, V., & German, A.J. (2016). ‘Can you estimate body composition in dogs from photographs?’, BMC Veterinary Research, 12(1). https://doi.org/10.1186/ s12917-016-0642-7

Eastland-Jones, R.C., German, A.J., Holden, S.L., Biourge, V., & Pickavance, L.C. (2014). Owner misperception of canine body condition persists despite use of a body condition score chart. Journal of nutritional science, 3, e45. https://doi.org/10.1017/jns.2014.25

German, A.J., Woods, G.R.T., Holden, S.L., Brennan, L., & Burke, C. (2018). Dangerous trends in pet obesity. The Veterinary record, 182(1), 25. https://doi.org/10.1136/vr.k2

Back to Blog