IMPROVEMENT OF NON-INVASIVE PREDICTION OF CARCASS YIELD IN FISH USING AN AI-DRIVEN PHENOTYPING PIPELINE COMBINING ULTRASOUND AND 2D/3D IMAGING
Introduction
Improving carcass yield through breeding program is essential to improve the economic and environmental sustainability of fish farming, particularly for species marketed as processed fillets, such as rainbow trout or salmon. Hence, predicting cacass yield accurately on breeding candidates is key to maximize the genetic gain achievable for this trait. The method currently used to predict carcass yield on candidates is to measure, using ultrasound, the thickness of the fillet below the ribs (E8) and the total depth of the belly cavity (E23) (Haffray et al., 2013). With those metrics, breeders are able to predict viscero-somatic index and thus carcass yield. The R2 between the ratio E8/E23 and carcass yield is about 0.3. The aim of the present study was to improve this prediction and to facilitate the phenotyping of carcass yield on candidates by developing a phenotyping machine combining ultrasound with 2D and 3D images analysed by an AI pipeline.
Materials and Methods
We developed an integrated phenotyping machine combining a 2D camera (Keyence VS Series), a 3D camera (Zivid 2+ MR60) and an ultrasound (IMV Exapad). In parallel, we developed an AI framework (FishFusionNet) capable to analyse the different information provided by these devices.
Prior to being processed by the FishFusionNet, the data from each device underwent specific preprocessing steps. For the 2D images, we developed two YOLO-based models: one to automatically detect 15 anatomical keypoints at strategic locations on the fish (FishKpt), and another to segment the different body parts (FishSeg). These segmentations were then used in combination with the 3D data to estimate volumes of the respective body parts. The 3D acquisitions also enabled the generation of a surface heat map of the fish.
Figure 1. Pipeline of the FishFusionNet AI model.
Finally, we developed YOLO-based models (EchoSeg) to segment E8 and E23 features on ultrasound images. To calibrate and validate the different models for rainbow trout, we phenotyped two different populations of fish derived respectively from Viviers de Sarrance (960 fish) and Bretagne Truite (1230 fish) nucleus.
Results and Discussion
The initial results of the FishFusionNet, trained on data from Viviers de Sarrance, show a substantial improvement of the prediction accuracy compared to the traditional ultrasound method, with an R2 of 0.62 between the carcass yield measured and the carcass yield predicted. However, these results do not yet include ultrasound data, as the EchoSeg model is currently under development. We therefore expect further gains in predictive performance with the integration of ultrasound features via EchoSeg, as well as through the inclusion of the data from the fish of Bretagne Truite. Then, since fish from both populations were genotyped, we will estimate the genomic correlation between the measured carcass yield and the predicted carcass yield. This is to verify whether predicted carcass yield represents the same genetic trait as the measured carcass yield.
Acknowledgment
These results are part of SEPIAA's project financed by France 2030 and Bpifrance.
References
Haffray, P., Bugeon, J., Rivard, Q., Quittet, B., Puyo, S., Allamelou, J. M., ... & Dupont-Nivet, M. (2013). Genetic parameters of in-vivo prediction of carcass, head and fillet yields by internal ultrasound and 2D external imagery in large rainbow trout (Oncorhynchus mykiss). Aquaculture, 410, 236-244.