New case study: using AI to improve gill health monitoring in Atlantic salmon

The application of AI for improved gill health diagnostics

The gills of Atlantic salmon are vital organs responsible for breathing, regulating salts and acids, and defending the fish against pathogens. In recent years, gill disorders such as proliferative gill disease (PGD) and complex gill disorder (CGD) have become more common, driven by warming seas, extreme weather, and new pathogens. These conditions are a growing challenge globally for salmon farming.

Gill health is currently assessed through histopathology - microscopic examination and scoring of gill tissue by trained specialists. While effective, this method is time‑consuming, subjective, and difficult to scale. Early, accurate diagnosis is essential for good fish welfare and effective farm management, highlighting the need for faster and more consistent diagnostic tools.

This project explored whether AI could reliably identify signs of gill disease in histopathology images. Using an existing dataset of gill tissue slides that had been manually scored by an expert fish pathologist, the team aimed to build a prototype automated classification system and create a demo diagnostic tool called GillsPipe. Valued at almost £80k, the project partners were University of Aberdeen, Scottish Fish Immunology Research Centre, Vertebrate Antibodies, Scottish Sea Farms, BioMar and Marine Scotland Science.

Researchers identified which tissue changes were the most important indicators of gill disease. Through a combination of statistical analysis, machine learning and expert review, three key markers were selected: lamellar hyperplasia, lamellar fusion, and vascular anomalies. These were used to train several AI models.

A Vision Transformer (VT) model - a deep‑learning approach - proved most effective. When trained on annotated images, it achieved high accuracy: 91%, 98%, and 97% for the three disease markers. The model produced both visual overlays highlighting affected areas and numerical scores indicating severity. Testing on new images showed strong agreement with expert assessments.

These results were used to build GillsPipe, which performed well in early trials despite being trained on a relatively small dataset. The project demonstrated clear proof‑of‑concept that AI can support rapid, objective, and scalable gill diagnostics.

Further development is needed, including larger datasets and computing capacity, but fully developed AI tools could transform gill health monitoring, support better welfare and survival, and strengthen productivity across the Scottish aquaculture sector.

The full title of this project is ‘Artificial Intelligence approaches to improve diagnosis of gill disease in Atlantic salmon (AIGD)’.

Read the case study