Computer Vision Detects an Association Between Gross Gill Score and Ventilation Rates in Farmed Atlantic Salmon (Salmo salar).

Abstract

Poor gill health compromises the health and welfare of farmed Atlantic salmon (Salmo salar) by causing respiratory distress and increased ventilation frequency. Poor gill health is caused by numerous factors, including amoebic gill disease (AGD), jellyfish stings, and toxic algae, and is monitored by fish farmers by manual 'gill scoring'. Gill scoring involves visual inspection of gill surfaces for visible lesions, such as white mucoid patches. In commercial salmon farming, these patches are commonly associated with AGD, a major cause of poor gill health. Manual monitoring of gills is labour-intensive, costly, and stressful for fish. This study tested a non-invasive computer vision approach to detect the association between the gross gill score and fish ventilation rates in commercial farms. We hypothesised that increased ventilation rates of farmed Atlantic salmon were associated with a higher gross gill score. The computer vision model first detected fish heads and classified their mouth states (open or closed) using a convolutional neural network, followed by a tracking-by-detection method to estimate ventilation rates by calculating the frequency with which fish opened and closed their mouths. Ventilation rates were estimated from 240 videos recorded at Tasmanian salmon farms and analysed alongside gross gill score, water temperature, dissolved oxygen, and fish weight. Multiple linear regression analysis revealed a positive association between ventilation rates and gross gill score, although the observed change in ventilation rates was relatively small. As laboratory diagnostic methods did not confirm AGD in this study, the gross gill scores should be interpreted primarily as indicators of gill health, acknowledging that they may also reflect signs consistent with AGD. While the tested computer vision method cannot serve as a diagnostic tool, it may assist the industry in identifying health and welfare issues that require further examination. This approach provides a non-invasive way to oversee health and welfare, enhances management practices, and guides manual health assessments.