A look at the susceptibility of osteochondrosis in swine through genetics.

by Øyvind Nordbø, PhD, Researcher, Topigs Norsvin

Osteochondrosis is a disease that is a major cause of lameness in pigs.

Susceptibility to osteochondrosis is partially influenced by genetics, which is why Topigs Norsvin addresses susceptibility to osteochondrosis through the breeding program by scoring osteochondrosis lesions from the high-definition CT (computed tomography) image data.

Scores are assigned by trained technicians using a subjective, 5-point scoring system.

Research is currently underway to update this process by replacing visual appraisal and scoring with an automated detection, and scoring pipeline with the help of artificial intelligence (AI) models.

Genetic Selection for Reduced Susceptibility to Osteochondrosis

Topigs Norsvin has selected against susceptibility to osteochondrosis since 2012.

Currently, osteochondrosis is scored by trained technicians based on visual observation of CT images. Osteochondrosis is scored for the lateral and medial condyle of both the elbow and knee joints, for a total of eight separate scores.

Assigning scores consists of the following steps:

1. Identifying the joint or condyle of interest (from a 3D image);
2. Assessing the severity of the lesion, and
3. Recording a score.

Scores are assigned at each individual location using a subjective, 5-point scoring system, then summed across all eight regions to calculate a combined score, ranging from 0 to 32. An individual’s genetic merit for susceptibility to osteochondrosis is estimated based on this combined score.

Selection for a lower combined score has proven to be effective, resulting in substantial genetic and phenotypic improvement in this trait.

Figure 1. Ten-year phenotypic trend for combined osteochondrosis score in the Norsvin Landrace line.

For example, phenotypic trends reveal a 70% reduction in overall osteochondrosis score in the Norsvin Landrace line throughout the last decade (Figure 1).

The next steps include improving the phenotyping pipeline by replacing the visual appraisal of osteochondrosis lesions with automated detection and scoring.

Segmentation

AI plays a significant role in medical image analysis, such as the identification and classification of pixels into different classes.

This process, referred to as segmentation, is the same method that Topigs Norsvin uses to classify various tissue types from CT image data.

Small Lesions in a Big Pig

One of the main challenges of developing artifical intelligence models for osteochondrosis detection is the size of the lesion relative to the full-body image.

For instance, a full-body 3D CT image consists of about 0.3 billion pixels, whereas the size of a large osteochondrosis lesion is only about 300 pixels. In general, the smaller the region of interest within an image, the more difficult it is to detect.

The best way to overcome this challenge is to focus on specific regions within the pig where osteochondrosis is most likely to occur.

The Anatomic Atlas

At Topigs Norsvin, the CT scanning process consists of collecting whole-body, 3D images for each pig. CT images are collected at market weight (approximately 286 lbs.) for all purebred boars at the Delta Canada and Delta Norway boar off-testing stations.

The first step in this process is sedation - sedating the pig enables technicians to load and position boars on the CT table as safely as possible.

During the scanning process, thousands of images are captured per individual.

These images contain pixels. While a given pixel may contain information on multiple tissue types, these tissues can be distinguished using AI models.

This same process was used to develop what we refer to as an AI-based anatomical “atlas.”

This provides an overview of the positioning and size of each anatomical structure within a CT image.

Figure 2. Automated segmentation of a whole-body CT image into 29 distinct tissue types, including bone, muscle, and organ tissue.

Using this atlas, the tissue of any CT image can be segmented into 29 different classes, including muscle, organ, and bone tissue (Figure 2).

Within the bone tissue, more specific algorithms are systematically used to identify joints connecting bones and any detectable instances of osteochondrosis lesions within these joints.

Joint Identification

Each major leg bone has two extreme points, which are located at the end of the bone. The center of a joint is identified based on the position of these extreme points of two neighboring bones.

Figure 3. Automated detection of the humerus (yellow), tibia and fibula (blue), extreme points (pink circles), and bounding box surrounding the stifle joint (pink cube).

Bounding boxes (Figure 3) are then constructed around the center point of each joint to define the region of interest for the identification of osteochondrosis lesions.

For example, bounding boxes improve the lesion-to-background ratio from approximately 1:1,000,000 to 1:10,000, reducing the computational complexity of segmentation in this region.

Looking at the Automated Segmentation of Osteochondrosis Lesions

In 2015, groundbreaking AI models were introduced for the segmentation of medical images (Ronnerberger et al., 2015). Since then, automated methods for the detection of diseased tissue have continued to develop, including state-of-the-art models for the segmentation of 3D images.

These models utilize information from all three dimensions simultaneously, which greatly improves the accuracy of segmentation, particularly for adjacent tissue types.

In an ongoing Norwegian research project, researchers from Topigs Norsvin, in collaboration with Associate Professor Kristin Olstad (Norwegian University of Life Sciences), are refining these models to segment lesions in the joints.

The first aim of this project is to build segmentation models to detect osteochondrosis in specific regions, including the shoulder, elbow, stifle, and hock.

The overall objective of this project is to combine these models with the current methods for joint detection to, ultimately, replace visual evaluation/manual scoring of osteochondrosis from CT images with a fully automated detection and scoring pipeline.

Once developed, this pipeline can be extended to detect and score osteochondrosis at even more locations within the body.

Topigs Norsvin’s CT Database

Topigs Norsvin maintains a database of CT image data collected on approximately 75,000 total animals. This database continues to grow at the rate of 10,000 additional animals per year.

This dataset is an extremely valuable reference population, used to develop/test new artificial intelligence models, novel traits, and, therefore, potential new selection targets.

Developing models to automatically detect and score osteochondrosis from CT images is a great example of how AI can be used in a modern breeding program to improve health, sustainability, and profitability in pig production.

---

Øyvind Nordbø
Øyvind Nordbø leads the Precision Phenotyping Research Platform for Topigs Norsvin. He has been working as a researcher with the company for the past 12 years, optimizing the use of genotype data in breeding programs and developing automated phenotyping based on data from various types of sensors, like CT, cameras, and electronic feeder stations. He holds a PhD in Biological Physics from the Norwegian University of Life Sciences.

---

---