AI now faster and better at finding lice larvae than human experts
Researchers used more than 120,000 images of sea lice larvae in sea water to train AI models to identify them. It means that a task that takes experienced biologists 30 hours can be done in 30 minutes by machine, with more accurate results.
It has always plagued and parasitised wild salmonids, but according to researchers, the farming boom has given Lepeophtheirus salmonis fantastic living conditions.
Despite years of intense work, research, prevention, measures and regulations, lice are still a burden on wild fish and one of the salmon industry's biggest problems.
New and accurate model
Now, researchers from the Norwegian University of Science and Technology (NTNU) and Wageningen University in the Netherlands have developed a new method that can provide better control of the parasites, reports the research magazine Gemini.
Using real images of lice larvae in the sea, production of synthetic data (real data cut and pasted together in new ways to increase the amount of data) and artificial intelligence, they have developed large datasets that AI models can train on to recognise lice.
In a new study, researchers show the method and how it can make lice hunting much more effective.
Better than experienced biologists
The latest study shows that trained biologists spent more than 30 hours over several days to identify 82% of the lice larvae in one large and complex sample of seawater.
The AI model the researchers developed needed 30 minutes to identify 97.5% of the larvae in the same sample.
“Many measures are being used and tested to combat sea lice. It is often the combined effect of several measures that will provide better health for both farmed salmon and wild salmonids,” says researcher Lars Christian Gansel.
He heads the Department of Biological Sciences at NTNU in Ålesund and has been involved in developing the method.
“To document the effect of methods in use, as well as the development and adaptation of new measures, we need more information about the spread of sea lice larvae. Our model makes it possible to obtain this information,” he says.
Traffic lights – for the sake of wild fish
Every year, between 400 and 450 million salmon smolts and rainbow trout are stocked in marine pens in Norway's fjords.
One fish farm can contain millions of salmon, and researchers claim this can spread millions of lice larvae into the fjords every single day.
The industry is regulated through the traffic light scheme which is intended to protect wild fish. If an area has a high lice count, salmon production must be reduced.
"If we are to succeed in eliminating the lice, it is best to prevent and prevent contact between the parasite and the fish. In order to develop, evaluate and document the effect of prevention methodsS, it is important to detect the larvae while they are still floating around in the sea," explains Gansel.
Difficult to count sea lice
It's not just sea lice larvae floating around Norwegian fjords. There can be as many as hundreds of thousands or millions of other organisms per sea lice larva in the sea.
In relation to the total amount of plankton and other particles, the louse can actually be described as a rare organism, according to Gansel.
"That's why we need to analyse large volumes of water to monitor sea lice. If we use too little water, we can easily overestimate or underestimate the number," he says.
Much has been tried to accurately count lice larvae, but with the methods that have been used for continuous monitoring of salmon lice larvae to date, the work is cumbersome, imprecise, time-consuming and expensive.
Created 120,000 lice pictures
“Available camera systems for analysing plankton often have too little resolution to distinguish species and developmental stages. There is still no fully documented method for continuous monitoring of sea lice in the sea,” emphasises Gansel.
Artificial intelligence and machine learning have emerged as an opportunity. The challenge has been a lack of high-resolution, clear, descriptive images of larvae in real seawater environments to train the AI models on.
Researchers at NTNU and Wageningen University may have found a solution. They have created their own video microscope and more than 120,000 images enriched with clippings of louse larvae and other organisms. They have trained AI models on the synthetic data.
“The models worked just as well as the experts using microscopes. Even though some other species in the sea may be similar, the model was able to distinguish what were sea lice in large real seawater samples,” says Gansel.
Hatched sea lice for AI training
The researchers have concentrated particles the size of sea lice from hundreds of seawater samples. In total, they have collected and filtered several thousand cubic meters of seawater at fish farms and sea areas near Ålesund.
Conditions in the sea can vary with seasons and locations. When there are few sea lice in circulation, it takes a long time to create good datasets from the samples.
To create more training material, the researchers hatched sea lice larvae themselves, added them to the water samples, and then let the water flow slowly through a glass tube while filming the particles in the water flow with a video microscope.
Cut, copy, rotate
Using a programme that can track and select individual parts in video, they separated images of larvae at two stages: newly hatched lice or nauplii, and the slightly larger copepodites that are ready to attach to the fish.
"The individual frames from the videos will not show all sides of the larvae. Maybe they are moving, and maybe they are just drifting in certain parts of the tube. Because we only see a few of all possible states, we can improve the models by creating synthetic data that is used in conjunction with many regular videos.
"Sea lice can vary slightly in size. To take into account the differences, we can scale the lice, turn them over, and include multiple lice in the same image. We can do the same with plankton and organisms that can resemble sea lice, to make the model even better," explains Gansel.
Removes a lot of uncertainty
The models can be used to monitor lice loads in areas where wild fish are expected. They can be used to calculate larval release rates and investigate how they disperse, grow and develop.
Monitoring can be a support for assessing possible measures, and is important when estimating the chance of infection between farmed and wild salmonids.
"Measuring larvae directly in the sea will remove some of the uncertainty in the current system, where the amount of larvae is calculated from the number of lice on the farmed fish. This can make the salmon lice map much more precise. Production can be planned better, and we can make better decisions about where we can farm, and what measures we can take against lice," says Gansel.
