Feedstuffs in Focus
Feedstuffs in Focus is a weekly look at the hot issues in the livestock, poultry, grain and feed industries. Join us as we talk with industry influencers, experts and leaders about trends and more. Feedstuffs in Focus is produced by the team at Feedstuffs.
Feedstuffs in Focus
AI for early poultry disease detection
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
In a poultry barn, disease does not wait for the next walk-through. We sit down with Dr. Guoming Li of the University of Georgia to talk about a practical question every grower and integrator faces: how do you catch health problems early enough to protect animal welfare, reduce losses, and safeguard food safety when time and labor are limited?
We explore precision poultry farming tools that turn everyday signals into early warnings. Dr. Li breaks down how thermography and machine learning can detect temperature shifts linked to avian influenza and Newcastle disease, and why focusing on non-feathered regions like the head and legs improves accuracy by reducing ambient-temperature noise. We also discuss how image-based diagnostics can fit into real farm routines, including the idea of a smartphone app that uses deep learning and transfer learning to classify fecal images for Salmonella risk assessment without adding expensive sensors.
Then we tackle the hard part: trust. When a model trained on one dataset fails on another region or housing system, it exposes the generalizability problem that still holds back AI disease detection. We also look at behavioral analytics, including the broiler activity index and computer vision tracking of movement patterns, as biomarkers for illness, stress, and abnormal conditions. Finally, we zoom out to what makes AI reliable in animal health: curated datasets, rigorous validation, and science-based inference instead of confident guesses.
If you care about poultry health monitoring, biosecurity, and practical AI on farms, listen now, share this with a colleague, and leave a review so more people can find the show. What signal do you think will become the most trusted early-warning tool: heat, images, or behavior?
Why Poultry Diseases Still Surprise Us
SPEAKER_00Infectious diseases such as avian influence and new chemical and phenomenella remain persistent threats to poultry health and food safety. The good news is that recent advances in artificial intelligence now allow for early on-harmed detection through image-based, thermographic, and behavioral analytics. Welcome to feed adopted and focus, our podcast, taking a look at the big issues affecting the livestock of poultry grain and animal feed industries. I'm your host, Jeremyhead. This episode of Feeds Inducted and Focus is sponsored by United Animal Health, a leader in animal health and nutrition. You can learn more about United Animal Health and how they're working to advance animal science worldwide by visiting their website at unitedanh.com. Joining RNH to discuss several machine learning studies aimed at advancing precision poultry disease diagnostics is Dr. Gingman Lee of the University of Georgia.
SPEAKER_01So Dr. Lee, before we get into the research, walk us through why we keep seeing these disease challenges across the industry. Is detection not soon enough or are there management areas we should be focusing on?
SPEAKER_02Yeah, I I think the poultry productions, I mean the characteristics, it may it just made the disease is consistently appear in the production system. Like you get a 20 to 30,000 birds in a boiler house, and then you get like 200,000 birds in a layer, in a cage layer system. And then within this densely distributed housing system, it is not inevitable to have disease or those diseases within the house will be um uh quickly spread within the house. And then it's not like we are not diligent to detect the disease, is it just short of hands or or in a routine manner? I mean it's just single person, it's an inspectors work through the house and then think about he has to check 20 to 30 birds in the house. It is very hard to find those birds. I mean, if he gets a problem, very obvious problem, he will call the veterinarian and go and check. And and they have to check the farms twice a day, at least twice a day, the contract growers. However, it's it is too hard. I mean you cannot take care of every bird in a house. And especially um and and one of the uh graph I present in a keynote speech is the bird age distribution of the uh poultry and approducers. And majority of the people are over 35 um years old. And then fewer and fewer young workforce want to get into raising the chickens or kind of things. And then it just reduces the energy or or the frequency in tracking the farms, or it is multi-factors, um social factors or housing system is not a single factor issue.
Thermography For Flu And Newcastle
SPEAKER_01So many many challenges there, it's important to maybe grasp some of this technology. You you talked uh quite a bit about it in your in your presentation. How does thermography combined with machine learning improve early detection of avian diseases such as um AI and um in Newcastle?
SPEAKER_02So so for the um avian influenza, especially uh hype highly pathogenic avian influenza, so once birds get it, it is very hard to correct it. Um and the birds will die very quick, and the symptoms will be very obvious, and then the farmers can easily see it because you have a lot of flaw, flaw ease and and and and that, I mean, you can check it. However, for the low pathogenic AB and funza, I mean if birds they they get affected, or new castle disease, uh, if birds uh get affected, so they will the body temperature will change. So they will have fevers, so those body temperatures will be higher than their normal rank. And then um in the thermographics kind of capture the surface temperature of a uh a subject. And then if you capture the birth temperature and then the temperature is raising high, and then we got those features for to machine learning out with and it can classify the disease positive or negative. Yeah, that is how they go.
SPEAKER_01Better than meat CI, that's for sure. Um why is focusing on non-feathered regions such as the head and legs, why is that critical for detecting early changes in broiler chickens?
SPEAKER_02Yeah, so I'm gonna give some backgrounds uh for these questions. So we actually conduct two set of studies. So one set is to uh classify the whole body profile temperature, and then we got uh successful detection um within 24 hours after infection. Um however if we focus on the head and the leg, which will have moniker area exposed, and then we got uh over 90 uh accuracy um of the classification within eight hours of uh infection. And then the reason behind it is like the thermographic, it can just capture the surface temperature, it cannot like penetrate through the feather. And then the cone body temperature it may near the skin, not near the feather. The feather may be subject to the environment or ambient temperature conditions. So if you focus on monaiche area like the head, you have eyes, or you have some skins exposed around the eye or the leg, you have some skin disposed, and then it can reflect the real temperature of the bird instead of the ambient temperature. And this will support the uh the classification.
SPEAKER_01That makes sense. Those feathers might be packing in more heat per pants than what the animal is actually.
Phone Photos For Salmonella Risk
SPEAKER_02Yeah, it's just the ambient temperature and the the bird home body temperature, the the differences. So we want to avoid the interference with that.
SPEAKER_01So, how does deep learning and transfer learning apply to classify fecal images for salmonella risk assessment?
When Models Fail In New Regions
SPEAKER_02So we uh the initial idea for the fecal image classification is like um the farmers they they have cell phones and they they go to the farm every day, and then rather than bring uh additional sensors, and then we think okay, maybe develop an app and then install on a cell phone will be an easy way for the farmer. And then they don't have to carry a lot of things uh when they inspect the chicken because they already have a very heavy load and they don't want to buy expensive equipment. And then if you have the app ready, and then they can take the fecal samples and then just fit to the machine learning classifier or deep learning classifiers, and then it can output the classification result, that would be perfect. And then the deep learning is a bit different from the conventional machine learning algorithm. It has very deep layers, and then that can go that can evolve features deeper and really unravel the data patterns. And then transfer learning is kind of so you have a model that already trains on one data set, and you don't want extensive efforts to train again from the beginning. I just want I just want the model already trained on a very big data set, but I just want to adjust a little bit in my specific application scenario, and then I can use the transfer learning, transfer the previous train weight, and then just adjust the prediction hat of the model. That can reduce the development effort for the disease classification.
SPEAKER_01What challenges did you encounter when validating the hybrid architecture across data sets from different regions such as America and in Africa?
SPEAKER_02Yeah, so that's a very good question. And then we uh when we developed the Samonala uh classification model, um we actually have two data sets. One was collected from African farm, African farm flow rating system. The other one was for uh the African data set. That was collected from the the cage system. Um the the manual drop on the conveyor belt for the African. So they are different. And then we want to test okay whether the model trained on American data set could be accurately detect the disease on uh African farm or opposite. And then it turns out it failed. And then um this truly indicates uh a generalizability issues of the model. So when that is the purpose for the cross-validation. So it's not so most of the models they train and test on the same scenario. Like you train on American data and you test on American data, you perform well. However, to test the full potential of the model, you're gonna cross-validate it. And then if you test it, it has a lot of problems in the generalizability.
SPEAKER_01So, how does the broider activity index serve as an early biomarker for infection and and what role do behavioral analytics play in disease detection?
SPEAKER_02So the the boiler activity index is kind of uh a very simple and easy uh techniques, uh, the the computer vision techniques, and then you have so for videos you have continuous or consecutive frames. And then the technique is kind of you subtract the consecutive frames, you get the difference between consecutive frames, and you summarize the pixels of the difference. And then the hypothesis behind this is like uh high levels of bone movement could result in high levels of uh active uh pixel changes and it will result in higher activity index. So this kind of thing is evaluate the glue levels or flock level activities. And then for instance, if a bird gets ill, or if a bird encounters heat stress, or if a bird is suffering from abnormal situations, like humans, we want to we want to rest, we want to rest. And then we want to reduce the activities, we want to reduce the movement. This can be directly reflected in the activity index. And then this can be captured automatically. And then we also explore the individual level differences. Like with the current computer vision algorithm, we can track the birds constantly, we can track the um the trajectories, uh the movement speed, and uh acceleration and movement distance, all these we call kinematic features can be extracted in the individual levels. And then like I mentioned previously, if an individual feel uncomfortable will reduce, or its trajectory is not performed as the normal birds. We test uh those kinematics for the birds with or without laminates, uh and the birds with laminates, they would have descripted or more concentrated trajectory than those birds with healthy birds. The healthy birds move the trajectories distributed everywhere within a pen. They show more continuous connected trajectory than those with laminates. So this is one example we test. And this is not only for laminates. We saw it can be scalable or it can be transferable to detect the disease, disease bursts. Yeah.
Generalizable Trustworthy AI And Data Quality
SPEAKER_01It's amazing what you can see in the barn using technology and that data. What are some next steps for you for further research and to validate some of these techniques for for broader application?
SPEAKER_02I think there are still a lot of things we have to adjust before we put the AI for the mobile or automatic uh poultry disease detection. The first one is the generalizability uh issues, as I mentioned. It tests on one, it tests uh on one data set and then it cannot perform well on the other ones. So we have to increase the variations or we have to adjust the model architecture to ensure the model is optimizing now uh to generalizable in different scenarios. And the other one is about the inference capability. So the model, I suggest that the current AI models, like like ChatGBD models, and then I put like uh seminar fecal images and ask, is it a causal dialysis infection? Ask different uh disease questions. And then the Chat GBD, I mean answer everything related to causal dialysis without thinking other things. So those current AI models kind of please the users. It's a user-driven platform. Instead of the science-based platform, it's a lack of inference be and provide science-based evidence answer. So this is something we have to improve in the future too.
SPEAKER_01We have to get that science-based information out there for it to catch up with us and where we're at.
SPEAKER_02Yeah, so you cannot always trust the general models. So you still have to uh get some specialized uh tools, and the tools have to learn very precise, trusty and reliable data sets. Because those general models they train a lot of things online. Those articles or those they can train on polls, they can I mean train on comments or they can train on any type of data. So those data is not quality control. And then garbage in, garbage out. And then you're gonna control the data for training the models. The data quality it is matter.
SPEAKER_01Absolutely. Dr. Lee, thank you so much for joining us today.
SPEAKER_02Thank you. Thank you for inviting me.
SPEAKER_00This episode of Feedstuffs in Focus was sponsored by United Animal Health, a leader in animal health and nutrition. You can learn more about United Animal Health and how they're working to advance animal science worldwide by visiting their website at unitedanh.com. I'm Sarah Muirhund, and you have been listening to Feedstuffs in Focus. If you would like to hear more conversations about some of the big issues affecting the Linux of Poultry Green and Animal Feed Industries, subscribe to visitpomodcondens.com, your favorite public condensed channel. Until next time, have a great day and thank you for listening.