Computer vision system using deep learning to predict rib and loin yield in the fish Colossoma macropomum

Area of Science:

• Aquaculture
• Animal Genetics
• Computer Vision

Background:

• Computer vision systems (CVS) offer non-invasive, low-cost phenotyping for aquaculture.
• Accurate measurement of economically important traits like rib and loin yield is challenging in tambaqui.
• CVS can predict morphometric measurements, aiding indirect selection in breeding goals.

Purpose of the Study:

• To validate a CVS using deep learning for automatic morphometric prediction in tambaqui.
• To estimate genetic parameters for growth and body yield traits in tambaqui.
• To assess the potential of CVS for large-scale phenotyping and genetic improvement.

Main Methods:

• A deep learning instance segmentation model (Mask R-CNN) was used for image analysis.
• Seven growth traits were measured, and fish were dissected into body regions (head, rib, loin).
• Pearson's correlation assessed agreement between manual and automated measurements; heritability and genetic correlations were estimated.

Main Results:

• High positive correlations were found between manual and CVS-predicted measurements.
• The CVS demonstrated high classification performance with minimal visible differences (approx. 3%).
• Heritability estimates for growth and body yield traits varied from low to high; favorable genetic correlations were observed for rib and loin yield.

Conclusions:

• The validated CVS is a resilient tool for large-scale tambaqui phenotyping.
• Rib and loin weight are under moderate genetic control, suitable for selection.
• Standard length and pelvis length can serve as effective indirect selection criteria for body yield.