Deep learning can automate chicken tibia-breaking strength quantification to improve animal welfare

TL;DR

A deep-learning system automates the assessment of chicken bone strength from X-rays, offering a faster and non-invasive alternative to manual methods.

Contribution

An end-to-end deep-learning pipeline is introduced for automated chicken tibia-breaking strength quantification with high accuracy.

Findings

The U-Net model achieved a Dice score of 0.91 for segmenting chicken tibiotarsus from X-rays.

The model's predictions correlated moderately (Pearson’s r = 0.74) with actual breaking strength measurements.

Predicted bone strength showed high genetic correlation with physical traits, supporting its use in breeding programs.

Abstract

Bone damage is an important welfare issue in the poultry industry, yet large-scale phenotyping of chicken bone strength currently relies on time-consuming manual annotation of X-rays or destructive post-mortem testing. To address this, an end-to-end deep-learning pipeline was developed that automatically (i) segments the chicken tibiotarsus from lateral X-ray images (U-Net, Dice = 0.91) and (ii) predicts its breaking strength from pixel intensities alone. Using 916 curated bone images, the predictor achieved moderately high correlation with measured breaking strength (maximum Pearson’s correlation of 0.74), exceeding the performance of a previous labour-intensive manual annotation method. Image-derived predictions were moderately heritable (h² ≈ 0.16) and exhibited an exceptionally high genetic correlation with the physical trait, indicating that selection on the model-derived phenotype is a good proxy to select for bone strength. The workflow therefore provides a potential rapid, non-invasive and genetically informative alternative to post-mortem testing, paving the way for the routine incorporation of bone-quality traits into commercial breeding programmes and improved poultry welfare at scale.

Figures