AQUA 2024

August 26 - 30, 2024

Copenhagen, Denmark

CATCH THE WAVE: A WORKFLOW TO IMPLEMENT IMAGES AND ARTIFICIAL INTELLIGENCE FOR SELECTIVE BREEDING

Y. Xue*, A.P. Palstra, H. Komen, R.J.W. Blonk, J.W.M. Bastiaansen

Department of Animal Breeding and Genomics, Wageningen University & Research, P.O. Box 338, 6700 AH Wageningen, the Netherlands.

yuanxu.xue@wur.nl

 



Introduction

Recent advancements in artificial intelligence (AI) have opened up new possibilities for measuring and monitoring breeding candidates by extensive data collection. Concurrently, automatic phenotyping using non-invasive computer vision techniques has gained significant attention. However, the abundance of methods and data requires a structured workflow to effectively address critical issues in aquaculture breeding. This study introduces a workflow (Fig. 1) specifically designed to maximize the benefits of integrating imaging and AI technologies into both new and existing breeding programs. We demonstrate the effectiveness of this workflow with a case study on critical swimming speed () of rainbow trout, a reliable indicator for swimming performance. We identified four swim traits genetically correlated to , and discussed the opportunities that these traits present for selective breeding.

Case study

In data collection, we captured 3D images of each breeding candidate after conducting swim tests to determine . To examine the physical characteristics influencing , we utilized a convolutional neural network or CNN-model trained on these images to predict  values. The result revealed that morphological variance in the fish accounted for only 12% of the variance in swimming performance. Visualization with Gradient-weighted Class Activation Maps (GradCAM) helped pinpoint image regions contributing to the predictions (see Fig. 2). With the interpretation of fish physiologists, we refined GradCAM into three specific morphological regions in the images with putative biological significance to . Four swim traits (see Table 1) were defined from these regions, and subsequently evaluated for their genetic correlations with . Our findings indicated that epaxial muscle volume showed the strongest genetic correlation (-0.48) with . Genetically, fish with larger and broader epaxial muscles, larger heads, and smaller caudal fins swim less fast. This suggests that increased fillet yield could be obtained by selecting for lower  in trout.