 And our next presentation is by Agnetta Olson, whose work is based out of Cardiff and Exeter University talking to mapping disease and distribution of fish using AI. My name is Agnetta Seymourson and I'm going to talk to you about our project, Computer Vision for Fish Disease Detection and Classification. So we know that many widespread fish diseases have visible symptoms that you can see with the naked eye. And Computer Vision is AI for understanding what's going on in still images and video. So we are hoping that we can use Computer Vision to supplement existing fish health monitoring using AI. So I am a computer scientist and physicist, but at the moment I'm doing a PhD in biosciences at Cardiff University. I'm supported by an excellent team of supervisors, also from biosciences at Cardiff University, Dr. Sarah Perkins, who's an expert on citizen science and on wildlife disease, Professor Joe Cable, who supports me on aquatic pathogens. Chris Jones is a professor at Cardiff University in the computer science department and he specialises in geographical information systems and retrieval. And Dr. Jack Christmas from Exeter University is an expert on computer vision. The research is funded via NERC. So I'm part of the GW4 Fresh Centre for Doctoral Training that they fund. We also collaborate with the Orkmore Angling Association who are helping us with a citizen science project that I'll outline in a bit. And we are working with the environment agency. So the aim of this project is to create a computer vision system that can detect and classify disease based on visible symptoms in still images of fish. And we also want to map the distribution of these wild fish and their health status. And we're starting with freshwater fish at the moment looking at Orkmore River but hoping to look at UK wide rivers. So the first step for us has been to collect images and look at their quality. Computer vision algorithms need lots and lots of examples of the classes that you want it to be able to classify into. So here on the left I've got an example of some dummy fish. Outlined in green are different examples of healthy fish and outlined in red are different examples of fish with some sort of visible symptoms of disease that could be fish eyes or sacralenia. And with all of these examples available the computer vision algorithm can learn what features define a healthy fish which ones define a fish with disease and how can it best separate the two so that the classification is as accurate as possible. Most of the images we've got at the moment are from online sources so angling forums, social media, anywhere online really. And also we have some good examples of fish with disease from the environment agency. We are working with these images on selecting the best algorithm for our problem. So we're looking to train algorithms which is what we call the learning process for computer vision algorithms. We've done a little pilot with limited data and we got 80% accuracy that was using the alex net neural net which is quite old now but still one of the first sort of good neural nets are available. Accuracy here means when how big the ratio is of the algorithm being able to predict what we have already labeled an image as so healthy or having sacralenia in this case. So this whole computer vision strand we want to combine with a citizen science trend. That's where we're working with the Ogmore Angling Association and we're hoping to have the pilot going live in October. They will be using iNaturalist which is an app where you can upload photos and then an AI in the background suggests a species of whatever you've uploaded a photo of and we can then download all of the images along with the species and with location and time information so that we can map the disease and just sort of fish distribution temporarily and spatially. So that's what we're doing at the moment. In the future we need to keep building our image database of freshwater fish with and without visible signs of disease because we need as many examples as possible. So if you're sitting on lots of images of freshwater fish and you think this sounds interesting then don't hesitate to contact me on the email below. In parallel with this we also need to look at the image quality and streamline the labelling which is the process of saying this fish is healthy this fish is not and there are lots of different software packages and different metrics for evaluating that are available and if anyone has recommendations or experiences I'd be very happy to discuss. Similarly with the algorithm selection and the training and the evaluation possible comparison with humans. If anyone has experiences they're willing to share then I would be very excited to discuss that as well. All right thank you very much for listening and I'm looking forward to answering any questions. Thank you very much Agnetta. So when I was listening to your talk it immediately jumped out to me that this is where we've spoken all day about different types of communities being involved and needing to cross over and here's an opportunity for us to cross over especially from FAO where I know there's a similar process happening with diseases of crops and trees and such like and I wondered if if you had had any opportunities to cross over with those groups who would also try to present or to develop workflows and maybe speak a little bit to this Alexnet and why you selected Alexnet. So for the first question we know that there's there's overlap with both looking at crops and looking at medical images as well. I don't have haven't had any contact with any groups just in my first year and yeah so any contacts on anyone who's doing similar stuff there would be much appreciated. When it comes to Alexnet and the choice of that it was just it's the first neural network that was shown to on this ImageNet dataset that people have been talking about previously. It performed really well when it's the first one that was a convolutional neural net that was showed that this was a possible thing to use for good quality classification. But as people have talked about the RCNNs and net REST 50 or by numbers there's lots of lots of different ones that you can choose depending on your computational resources and other things yeah. When we were starting to discuss how to run this event this initiative for the forums one of the sessions we had was to just get the other teams in here the PlantNet team the other types of plant diseases team and that was thought to be a very good way to go apart from the fact that we sort of ran out of time we have that many people would like to talk about what they do but maybe this is an opportunity for FAO to play a role of bringing those other groups to the table where we can alert people who've come to this or registered for this event just to allow them to see what's happening in these other domains. So thanks very much Agneter.