WEBVTT Kind: captions Language: en 00:00:00.400 --> 00:00:04.080 We will be hearing from Dominique Duncan   00:00:04.080 --> 00:00:09.200 at the University of Southern California  and Dominique we're ready whenever you are. 00:00:10.160 --> 00:00:18.080 Thank you. So I'm Dominique Duncan from USC  from the Nerve Imaging and Informatics Institute   00:00:18.080 --> 00:00:25.280 and I'll be talking about our COVID-19 data  archive. For short it's called COVID ARC. 00:00:26.320 --> 00:00:33.200 So we have a lot of experience at our institute  with large scale multimodal data repositories   00:00:33.760 --> 00:00:40.960 so we decided to apply our experience and  the tools that we've developed in these other   00:00:40.960 --> 00:00:48.400 projects and extend them to this COVID-19 data  archive. So you can see here on the bottom left,   00:00:48.400 --> 00:00:54.560 that's a screenshot of our home page. If you want  to go to the website, I encourage everyone to go   00:00:55.120 --> 00:01:03.360 to covid-arc.loni.usc.edu to find out more about  the project, the various data sets that we have,   00:01:03.920 --> 00:01:11.920 our analytic tools and to find out more about  either uploading data or downloading existing data   00:01:11.920 --> 00:01:19.600 that we have. So we have some publicly available  data sets that we've been curating and organizing.   00:01:20.560 --> 00:01:27.680 We want to encourage researchers to  be able to not only perform analysis   00:01:27.680 --> 00:01:35.600 on individual data sets but across data sets from  various sites and so we want to make that process   00:01:35.600 --> 00:01:43.120 easier for them. And so we're also providing  various tools like quality control tools   00:01:44.000 --> 00:01:51.360 that people can use to evaluate images and  evaluate the quality using various metrics. We   00:01:51.360 --> 00:01:58.480 have various visualization tools for imaging data  and other types of data and then a wide variety of   00:01:58.480 --> 00:02:03.840 analytic tools that people can use. And then  for data sets that are not publicly available,   00:02:04.880 --> 00:02:13.120 data providers can decide if they want to store  their data on our server or if they want to keep   00:02:13.120 --> 00:02:19.840 their data stored locally at their site and just  provide us with the metadata so that we can let   00:02:19.840 --> 00:02:24.640 users know what data are available and  then we just facilitate the process   00:02:25.200 --> 00:02:32.080 of requesting access to that data. So if data  providers give us their data, it's anonymized   00:02:32.720 --> 00:02:39.600 and the data providers maintain full control of  access and they get to decide who gets access to   00:02:39.600 --> 00:02:45.520 it. If they want to wait to publish some findings  and then make their data sets publicly available,   00:02:45.520 --> 00:02:52.400 they're able to do that as well. And then  we use ASPERA which is IBM's HIPAA compliant   00:02:52.400 --> 00:02:59.200 encrypted high-speed file transfer system  for either uploading data to our server or   00:02:59.200 --> 00:03:03.840 for people who want access to the data.  They can download it that way as well. 00:03:05.440 --> 00:03:09.760 I know you can't see any of this text  but this just gives you an overview   00:03:09.760 --> 00:03:16.560 of what we have on our server. So, if you  request access, this is the tree structure   00:03:16.560 --> 00:03:22.960 of the COVID ARC project and so it's separated  into the data and then our analysis that we've   00:03:22.960 --> 00:03:29.440 been working on because we're also very involved  in the analytics for this project. So, the data   00:03:29.440 --> 00:03:36.000 will be separated by the different sites and  then there's more information on each of those. 00:03:36.000 --> 00:03:41.600 So, here is a table of the data sets  that we currently have on the server   00:03:41.600 --> 00:03:49.200 so you can see the location, where the data were  acquired for each of those data sets, and then   00:03:49.200 --> 00:03:55.680 the data format of those images. Right now,  we're focusing mainly on chest CT but we're   00:03:56.640 --> 00:04:03.520 also interested in other types of data and  soon we'll have some brain data so brain   00:04:03.520 --> 00:04:10.560 MRI of COVID-19 patients as well as  EEG. So here you can just see how many   00:04:11.600 --> 00:04:17.600 COVID images and non-COVID images and  masks are there from each of those sites. 00:04:18.960 --> 00:04:21.520 Here you can see some class activation maps.   00:04:22.480 --> 00:04:29.200 So there are many features, as we know,  CT features of COVID-19 patients like   00:04:29.200 --> 00:04:34.880 ground glass opacity, consolidation, crazy  paving pattern, reticular pattern, etc.   00:04:35.680 --> 00:04:42.080 So this just highlights the regions that were most  important for our classification that we're doing. 00:04:43.280 --> 00:04:50.320 One of the issues that we're discovering  is that image quality tends to   00:04:50.320 --> 00:04:58.320 lead to misclassification and so we're  assessing the quality of the images but then   00:04:58.320 --> 00:05:02.640 also considering various ways of  improving the quality of the images   00:05:03.200 --> 00:05:08.320 so looking at different filtering methods and  that's something that's currently in progress. 00:05:09.680 --> 00:05:13.200 We're also doing image  thresholding on the lung masks   00:05:13.200 --> 00:05:17.280 so we're using image thresholding  to find the best possible   00:05:17.280 --> 00:05:24.240 processed image type in the lung masks to improve  the prediction rate of our neural networks. 00:05:24.240 --> 00:05:29.440 And this is a table on one of the  data sets. That first one from Brazil   00:05:29.440 --> 00:05:36.960 that had about 1200 COVID patients and 1200  non-COVID images and we did a comparison   00:05:38.080 --> 00:05:42.800 looking at various methods to  compare convolutional neural networks   00:05:42.800 --> 00:05:48.000 and their accuracy. So we found that  ResNet-18 performed the best for that. 00:05:49.760 --> 00:05:54.080 And just to summarize because I think I'm  at the end of the time these are the people   00:05:54.080 --> 00:06:00.800 that are working on the project including the  REU students. And I'd like to thank the NSF for   00:06:00.800 --> 00:06:07.360 funding this project as well as Katie and Florence  for organizing this and inviting me to talk. And   00:06:07.360 --> 00:06:14.240 please visit the website or email me if you have  any questions and if you would like access to   00:06:14.240 --> 00:06:19.840 any of the data or if you have some data that you  would like to contribute to the website. Thanks.