Computer vision-based morphometric models developed under the DigiHeart project allow the quantification of shape variations in salmon hearts. Understanding the link between heart shape and mortality can help to manage hatching, production, and environmental conditions to improve fish welfare and reduce fish mortality.
For aquaculture to be ethically acceptable and sustainable, as well as to protect the investment made, mortality rates need to be lowered. Currently, around 20% of farmed salmonids die at sea before slaughter.
Death from infectious heart diseases such as cardiomyopathy syndrome (CMS) has been studied in detail, while death from non-infectious heart diseases has received little attention. The shape of farmed salmon hearts is different from that of wild salmon, and this may relate to fish health and performance.
Abnormally shaped hearts could cause cardiac arrest under stressful conditions, such as handling and parasite treatment. Farmers lack models and tools to fully understand the causal relations between the production parameters, heart shape, function, and mortality, as well as predicting the outcome of stressful interventions.
Computer vision-based morphometric models
The DigiHeart project has collaborators from both academia and industry in Norway, Sweden, and the Faroe Islands. Working with the project partners, DNV has taken advantage of the knowledge regarding different production practices and the prevalence of heart diseases in Norway and the Faroe Islands to analyse and understand the causes behind this critical issue.
A key deliverable of the project has been the development of computer vision models and machine learning to detect anomalies in fish (Atlantic salmon) hearts. Using the concept of morphometric analysis, it was possible to quantify the shape of fish hearts by identifying key points of interest on the hearts and calculating the distances between them. The morphometric scores obtained in this way were used to classify heart images. The algorithm was able to successfully differentiate between the three different projections (angles) at which the heart photos were taken, and also identified ventricle elongation as a key distinguishing feature.
The benefits
Researchers have identified 38 visual characteristics that define the shape of a salmon heart. Manual classification is a laborious and subjective task. DNV researchers automated the identification of key heart shape features and classification using machine learning. Models can be further developed to understand the causal link between shape and mortality.
Furthermore, by identifying operational practices and environmental factors that underlie the development of heart shape variations and health, it becomes possible to mitigate production losses in aquaculture. The knowledge derived from this project will also assist in optimizing aquaculture practices to improve fish welfare and reduce mortality rates.The three-year collaborative project is NordForsk-funded and ends in December 2023.
