A technology that uses deep learning to predict, with a high degree of accuracy, whether the eggs of the Pacific bluefin tuna will hatch after being spawned has been developed by a joint research team including Associate Professor Kei Terayama of the Yokohama City University Graduate School of Medical Life Science, Project Assistant Professor Naoto Ienaga of the Keio University Graduate School of Science and Technology, Chief Researcher Kentaro Higuchi, Group Director Toshinori Takashi, and Manager Koichiro Gen of the Aquaculture Division of the Japan Fisheries Research and Education Agency (FRA), and team leader Koji Tsuda of the RIKEN Center for Advanced Intelligence Project. This system is expected to optimize the process of selectively hatching and raising high quality eggs for the production of bluefin tuna seed (fry for fish farming). A report on the development was published in Scientific Reports.
Image showing eggs immediately after spawning. Egg viability is determined from cytoplasm and egg contour.
As a well-known option on sushi restaurant menus, Bluefin tuna is one of the leading elements of Japanese food culture, but in recent years the number of surviving tuna has reached the lowest levels recorded, making the sustainability of the resource a major issue for the continued consumption thereof. To that end, expectations are rising for technology to mass produce bluefin tuna fry using completely aquaculture-based methods, independent of natural resources, and a shift to fish farming using the technology.
Kindai University was the first in the world to achieve the production of bluefin tuna completely through aquaculture in 2002, and the FRA has been working on artificially controlling bluefin egg production using massive land-based tanks. However, the survival rate of the farmed bluefin is still drastically low compared to other farmed fish such as red seabream or salmon, so there has been a need for the development of a more efficient seed production technology.
One important challenge in fish seed production is increasing the efficiency thereof by evaluating egg quality to predict whether eggs will hatch normally and survive. This is because the ability to increase egg quality will enable the survival of a greater number of fry, and the use of egg quality prediction soon after spawning may enable the fry to be raised more efficiently.
Unfortunately, adequate progress has not yet been made in research on the prediction of bluefin egg quality and the characteristics required for that prediction. To date, egg appearance (including egg size and the shape and quantity of oil particles) has typically been used for predicting and determining egg quality, but research has been inadequate because many of the shapes and other characteristics involve subjective elements that are difficult to verbalize or quantify.
Accordingly, the research group collected 290 bluefin tuna eggs immediately after spawning at the FRA, photographed them via microscope, and tested the hatching thereof. Three images were obtained for each egg, focusing on cytoplasm, egg contour, and oil particles. Furthermore, data collected on hatching included whether each egg hatched normally, and whether the period of time each fry survived without food after hatching was less than 4 days or more than 5 days.
The team then built the egg quality prediction system. The system uses a deep learning shape detection algorithm called Faster R-CNN to extract only egg images from the images collected via microscope. A deep learning neural network called VGG16 is then used to predict whether each egg will hatch normally and whether it will survive for 5 or more days. The network was then trained using the egg images and hatching data.
An analysis of the prediction accuracy using 10-fold cross-validation showed an F score (an indicator where the prediction accuracy is higher the closer the number is to 1) of 0.911 for predicting normal hatching, and 0.875 for prediction survival without food beyond hatching. In each case, it was found that highly accurate prediction was possible. Prediction accuracy was found to be particularly high when focusing on egg contour and cytoplasm. The prediction accuracy was in fact higher than that of the four experienced aquaculture researchers on the team, demonstrating the effectiveness of egg quality prediction using deep learning.
When a method called Grad CAM was used to visualize which portion of an image contributes most to prediction outcome, it was found that there was particular focus on cytoplasm and egg contour, including on portions where cell shape seemed to be slightly deformed. This was in alignment with the portions considered important by the four expert researchers, which can be interpreted to mean that the deep learning neural network successfully identified the important portions for egg quality prediction.
The expectation is that more efficient fry production will be possible if batches with a high number of eggs predicted to be high quality are prioritized for cultivation using the system the team developed. Furthermore, this approach to egg prediction applied to bluefin tuna may also be applicable to use in the farming of other fish varieties for improved accuracy as well.