Teaching the computer to recognise suboptimal skeletal growth/osteochondrosis, to use this in selection for sustainable legs

Behind a long and complicated title, there is really a rather simple aim: we wish to use artificial intelligence to improve evaluation of skeletal health in pigs. Norway is a leading country in the use of computed tomography (CT) in pigs Having healthy and sound pigs depends just as much on which parents they are bred from, as on how they are raised. Selecting which pigs to breed from is therefore an elaborate process, especially for the boars who become parents to considerably more piglets than the sows. Some traits are measured in live pigs, whereas others are assessed after slaughter. For breeding boars, the latter traits historically had to be extrapolated from slaughtered relatives. However, in 2008, Norsvin, that is: Norwegian pig breeders, purchased a CT-scanner (also known as a CAT-scanner) and since then, it has been possible to scan an entire, sedated boar in 90 seconds. Once the scan is done, the pig can be cut up into virtual pieces, without having to use a knife. This is important because it means traits can be measured in the boar itself, in addition to its relatives, and other countries are following suit. Artificial intelligence can be used to find things in CT-images Most people who own a smartphone are aware that it recognises faces, for example when taking pictures. This is a form of artificial intelligence based on so-called machine learning: if you label something in an image and tell the computer enough times what it is, then it can, via networks that resemble the brain, learn to recognise it by itself. Osteochondrosis is a developmental disease that is common in the skeleton of pigs, horses, dogs, humans and other species. The disease occurs due to a failure in blood supply leading to small infarcts in growth cartilage, and as the animal grows, the infarcts become visible as characteristic notches in the bones. It should be possible for artificial intelligence to learn to recognise such notches. In the project, the computer has learnt recognise osteochondrosis by artificial intelligence In the project, we have focussed on four joints in CT scans of 201 pigs: the shoulder, elbow, stifle/knee and hock/ankle. According to the original aim, we also investigated osteochondrosis in growth plates and the spine, but could not find any evidence that it causes problems there in young pigs, so this was set aside for now. The four mentioned joints were cut out from the CT scans so they could be straightened out, because pigs often lie with their legs crossed in the scanner. Afterwards, an expert radiologist drew around osteochondrosis-notches in the four joints with different colours, pixel by pixel. This is as simple as it sounds, but it takes awfully long time. The coloured areas were fed into a computer that looked at them again and again in multiple cycles, to learn to recognise them. This was an exciting time where the computer was started in the afternoon, kept learning through the night, and then one could get up the next morning to see how clever it had become. The more cycles it does, the better it learns, and after a considerable amount of time, we are happy to say that the computer has learned to recognise osteochondrosis in four joints! The results support that machine learning and CT will be even more important in future The machine learning performs better in the stifle/knee than in the other three joints, readily explained by the fact that there were more lesions to practice on in the stifle. Historically, osteochondrosis has been registered in a way that can be described as counting affected regions by hand. With machine learning, it is now possible to push a button, and the computer lists the exact number of lesions and their size in cubic millimetres. The project has therefore resulted in development of a quick and powerful tool that we are excited about using in future research, breeding and keeping of pigs. The tool means that labour can be freed up from counting lesions, to making more informed and therefore better evaluations about how to use the results to achieve the best possible health and welfare for the maximum number of pigs. The project has already given us some new information in that respect: it appears that there is a relationship between pigs having many lesions also having small lesions, and small lesions are less dangerous because they tend to resolve by themselves. In comparison, pigs with few lesions can have either small or large lesions, and large lesions progress more readily to serious disease. If one only counts lesions, there is a risk of unknowingly selecting for more serious disease, and the project therefore supports that both number and size should be considered when selecting the best pigs to breed from. It is easier to measure size in CT images than by any other method, so this information supports both how important CT can be, and that it is extremely useful to be able to use machine learning to find things for us in images.

Prosjektet har hatt den virkningen at det har gitt oss et nytt verktøy i verktøykassen: vi kan lese av osteochondrose automatisk i fire ledd. I øyeblikket må leddene klippes ut fra CT-bildene først, men det gjøres også automatisk. For forsker-brukerne betyr dette endret praksis fordi man kan bruke verktøyet direkte i forskning der man har ressurser og tid til å klippe ut leddene før avlesning. For industri-målgruppen gjenstår det fortsatt et valg mellom å bruke tid på å klippe ut leddene før avlesning, eller på å oversette avlesningen fra utklippede ledd til hele griser før implementering. Valget koker ned til prioritering av hva man skal bruke datakraft på, og må balanseres mot de automatiske avlesningene som gjøres allerede av kjøtt- og fettprosent m.m. Når valget er tatt forventes prosjektet å ha stor virkning for industrien i form av endret praksis fra manuell til automatisk avlesning av osteochondrose. Prosjektet har også hatt den virkningen at det har endret måten de deltagende fagmiljøene ser på hele verktøykassen. Osteochondrose er en sykdom som oppstår og gror, dvs. oppfører seg dynamisk så lenge dyret fortsatt vokser, slik som i dette prosjektet. For fagmiljøet som jobber hovedsakelig med programmering har det vært en utfordring å jobbe med input som er dynamisk, ikke statisk, men de har tatt den på strak arm. Programmeringsmiljøet har utnyttet eksisterende kompetanse maksimalt og ervervet ny kompetanse i form av forståelse og håndtering av dynamiske data som de vil dra nytte av i fremtiden. Det veterinærmedisinske fagmiljøet har virkelig blitt inspirert og fått øynene åpnet for at biologiske spørsmål man tidligere måtte svare på gjennom nitid manuell registrering og analyse ofte kan besvares gjennom litt programmering i MatLab. Samlet sett betyr dette at vi nå ser helt nye måter å stille spørsmål, registrere og analysere data på som vil påvirke planleggingen av alle fremtidige studier. Dette gjelder særlig for omfattende prosjekter som krever store datamengder; her er tverrfaglig samarbeid som det nåværende prosjektet veien fremover, og vi kommer til å samarbeide på denne måten igjen. Seleksjon mot osteochondrose er viktig fordi det reduserer tap forbundet med griser som ikke når sitt potensiale på grunn av uhelse. Prosjektet kan således bidra til bedre forvaltning og mer bærekraftig matproduksjon i fremtiden. Genetikk brukes allerede i seleksjon, men det er basert på kunnskap om hvilke regioner av genomet som er involvert, ikke nøyaktig hvilke gener eller hva de gjør i osteochondrose. Det er vanskelig å svare på fordi det dreier seg om mange gener. En vakker dag i nær fremtid kan man koble prosjektets forbedrede registrering med genetisk analyse for tusenvis av griser, og da er svaret innen rekkevidde. Når det kommer vil det ikke bare eliminere feilkilder i registrering av osteochondrose på gris; det vil også kunne ha betydning for andre arter inkludert mennesker.

The skeleton in the spine and limbs grows by endochondral ossification. The most important disease that affects this process is osteochondrosis, which can be heritably predisposed. Osteochondrosis in joints can give loose fragments (osteochondrosis dissecans); osteochondrosis in growth plates can give angular limb deformities, while osteochondrosis in the spine can give hunchback and other deformities. Loose fragments can progress to painful osteoarthritis. Angled limbs or spine can also lead to abnormal loading and predispose for osteoarthritis, or claw problems that lead to early loss of breeding sows. In sum, there are few other diseases that influence the health, welfare, performance and longevity of pigs to the same extent as osteochondrosis. It is therefore extremely important to prevent this disease through selective breeding. All Norwegian potential breeding boars undergo computed tomographic (CT) scanning for automated quantification of lean meat and fat percentage for selection. Osteochondrosis is also manually evaluated in eight different places, something which takes a relatively long time per pig. Meat and fat percentage can be collected based on grey values in the CT scan, but for osteochondrosis it is unfortunately not that simple. We therefore plan to use machine learning, or artificial intelligence, because the technique can solve more complex tasks. In this project, machine learning will consist of a veterinarian marking osteochondrosis lesions in a number of CT scans. Thereafter, the computer will compare the images with and without markings and use so-called neural networks to teach itself to make the same markings as the veterinarian in unlimited numbers of CT scans. According to the aims, the computer will first be taught to recognise osteochondrosis in joints, then in growth plates, and finally in the spine, before the three machine learning protocols are joined together to make one model for automated whole-body scoring of osteochondrosis.