Phenomics

Automated osteochondrosis detection from CT scanned pigs; genetic parameters and comparison with manual scoring

---

Abstract

The objective of this study was to develop and assess an automated computed tomography (CT)-based framework for phenotyping osteochondrosis (OC) in pig's legs. OC is a developmental disorder, representing a major cause of lameness in pigs. For the past 15 years, Topigs Norsvin has employed a manual CT-based protocol for scoring osteochondrosis in knee and elbow joints. With moderate heritabilities (0.2-0.4), a large reference population of ~50,000 animals with combined genotypes and phenotypes, and sufficient emphasis on the breeding goal, substantial genetic and phenotypic progress has been achieved. However, manual scoring is labor-intensive, creating a need for automated detection and quantification of osteochondrosis lesions. This task is challenging because lesion volumes are extremely small relative to the full CT volume of a 120 kg pig. We propose a multi-step framework for automated phenotyping. After segmenting major anatomical structures and identifying joint locations, a transformer-based 3D volumetric medical image segmentation model, nnFormer, was trained to segment OC lesions within localized bounding boxes. The OC segmentation model was trained on a set of 198 labelled pigs, whereof 20 animals constituted an independent test set. The lesion-segmentation model achieved an F1-score of 0.592. CT scans of 1,502 genotyped Landrace test boars were then run through the automated OC segmentation model. Lesion volumes across joints were aggregated into phenotypes representing two phenotypes: (i) elbow_and_knee, which covers lesions in those joints only, and (ii) all_joints, which represent the sum of lesion in shoulder, elbow, knee and hook in mm3. Finally, we conducted bivariate genetic evaluations based on manual scoring, in conjunction with automated volume-based phenotypes. Phenotypes from automated segmentations gave higher heritabilities than those human scoring: h2all_joints=0.38 (±0.04) and h2elbow_and_knee=0.44 (±0.04) compared to h2manual=0.26 (±0.04). Genetic correlations between automated and manual scoring were high, 0.98±0.03 for elbow_and_knee and 0.94±0.03 for all_joints. These results demonstrate the potential of automated frameworks to increase the efficiency and accuracy of phenotyping, ultimately improving breeding programs.

Keywords: 2026

How to Cite:

Hassan, M., Nordbø, Ø., Olstad, K. & Aasmundstad, T., (2026) “Automated osteochondrosis detection from CT scanned pigs; genetic parameters and comparison with manual scoring”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2286378. doi: https://doi.org/10.31274/wcgalp.23969

Rights: 1