WEBVTT Kind: captions Language: en 00:00:00.160 --> 00:00:01.360 Katie Naum: Next, we have   00:00:02.240 --> 00:00:07.200 Michela Pazzani and Albert Hsiao of the  University of California at San Diego   00:00:08.000 --> 00:00:13.040 who will be telling us about explaining machine  management for COVID-19 management, so I’ll let   00:00:13.040 --> 00:00:17.280 you guys take it away. Thank you. Michael Pazzani:  00:00:17.280 --> 00:00:22.320 Okay so first I'd like to thank NSF, particularly  information intelligence systems, for funding   00:00:22.320 --> 00:00:28.160 this, and this is a collaboration between myself,  a computer scientist and Albert, a radiologist   00:00:28.160 --> 00:00:34.080 with a deep background in deep learning, and we're  using machine learning methods, particularly deep   00:00:34.080 --> 00:00:41.280 learning, to analyze CTs or x-rays to manage  COVID-19. While a diagnosis is important, we're   00:00:41.280 --> 00:00:47.600 also concerned with understanding the severity  of this disease from the imaging, and finally   00:00:47.600 --> 00:00:54.800 it's not sufficient just to say you have a 96%  chance of having COVID but, or even to highlight   00:00:54.800 --> 00:01:00.320 a portion of the lower lung, but we aspire to  label the images with things like, there's ground   00:01:00.320 --> 00:01:06.960 glass in the lower left lung. We're evaluating  a variety of existing approaches for both   00:01:06.960 --> 00:01:13.360 classification and explanation and starting to  develop new ones as well. The ultimate goal is for   00:01:13.360 --> 00:01:19.120 machine learning to acquire diagnostic signs that  can be communicated to people, such as a clinician   00:01:19.120 --> 00:01:24.560 when they're doing diagnosis, or perhaps to  teach peers or residents without even [inaudible]   00:01:24.560 --> 00:01:32.720 computers marking up the images. On the next  slide we show the process of deep learning-  00:01:34.800 --> 00:01:40.800 -and essentially we have a database, an existing  database of images as well as new images that   00:01:40.800 --> 00:01:47.440 unfortunately we've collected at UCSD over the  past few months of patients with COVID, and the   00:01:47.440 --> 00:01:55.360 goal is to take images of normal patients, those  with COVID-19, those with other conditions and   00:01:55.360 --> 00:02:02.240 come up with a learning method that distinguishes  them, and also provides the explanation. We're   00:02:02.240 --> 00:02:06.800 quite fortunate in that we get to leverage a lot  of existing infrastructure that was already in   00:02:06.800 --> 00:02:14.080 place of doing imaging, sending it to the cloud  for analysis, and then sending it to the clinic   00:02:14.080 --> 00:02:19.600 where the physician end user can observe it.  All of this was in place already by Albert at   00:02:19.600 --> 00:02:27.760 UCSD Health and we've just had to modify the cloud  diagnostic procedures with new data from COVID-19. 00:02:27.760 --> 00:02:28.960 And from here I'll let   00:02:28.960 --> 00:02:32.080 Albert take over and explain a little  bit more about how we're doing this. 00:02:32.080 --> 00:02:36.000 Albert Hsiao: Thanks Mike, thank you again for giving us the   00:02:36.000 --> 00:02:42.960 opportunity to present this work. It's definitely  very technical in nature but also very clinical   00:02:42.960 --> 00:02:51.200 and impactful immediately as we're already using  it in our clinic. The primary concept is really to   00:02:51.200 --> 00:02:58.720 develop AI algorithms that allow us to localize  pneumonia. This is work that we started even   00:02:58.720 --> 00:03:07.840 before the COVID-19 pandemic began, but has  been accelerated a lot because of the need.   00:03:08.960 --> 00:03:15.920 One really key important aspect of COVID-19 is  that not every patient develops pneumonia, some   00:03:15.920 --> 00:03:20.800 patients do and some patients don't, some become  asymptomatic, of course, but those that do,   00:03:21.440 --> 00:03:29.840 the severity of pneumonia on x-ray or CT provides  us very good prognostic information and a lot of   00:03:29.840 --> 00:03:35.280 data is starting to come out with that. We've  taken a very different strategy towards this   00:03:35.280 --> 00:03:40.960 U-net type segmentation approach as opposed to a  lot of classification approaches that have been   00:03:40.960 --> 00:03:48.320 previously used, although both are feasible, and  you can generate these kinds of probability maps   00:03:48.320 --> 00:03:53.680 and what-not - activation maps from classification  approaches which we'll be exploring as well.   00:03:55.040 --> 00:04:02.080 The important aspect is quantifying the severity  of illness, I guess, essentially gives us   00:04:02.080 --> 00:04:07.440 prognostic information, because ultimately we want  to know which patients require hospitalization,   00:04:07.440 --> 00:04:12.880 which ones can stay at home, which ones require  mechanical ventilation and which ones are likely   00:04:12.880 --> 00:04:20.080 to survive or not. And some of our initial data  here is showing us that those patients with   00:04:20.080 --> 00:04:26.400 high likelihood predicted by the - by the  algorithm are also the ones that tend to   00:04:27.280 --> 00:04:34.720 not survive and are the ones that tend to require  intubation. So, this will give us really good   00:04:34.720 --> 00:04:39.440 data for how to best manage these patients  so that's a really critical element of how-  00:04:39.440 --> 00:04:43.680 -how this will all come into play in our  clinic and hopefully many others through   00:04:43.680 --> 00:04:49.280 our collaborations. Our current results in  COVID-19, this is one example of a patient   00:04:49.280 --> 00:04:56.160 with COVID-19 who presented to our clinic. Our  AI algorithm produced our initial AI algorithm   00:04:56.160 --> 00:05:01.040 produced this result, very subtle that doesn't  really highlight the areas of pneumonia that well   00:05:02.160 --> 00:05:09.040 as it was trained on initially only public data  before COVID-19 and we came up with a strategy   00:05:09.040 --> 00:05:13.840 that uses active learning transfer learning  to specifically identify good cases for us to   00:05:14.400 --> 00:05:21.520 train on applying transfer learning to that neural  network. Using also concurrently performed CTs   00:05:21.520 --> 00:05:27.840 that were done to give us a better ground truth  and that that's given us a higher performance   00:05:27.840 --> 00:05:31.520 both on the internal and external data set  and really highlights the pneumonia better so   00:05:31.520 --> 00:05:36.320 our initial algorithms that were in place  we're replacing with these updated algorithms-  00:05:37.040 --> 00:05:45.760 -thanks to the support of the NSF in this project  we've been deploying this it's in our clinic and   00:05:45.760 --> 00:05:52.880 there's certainly articles online about it as well  as a peer-reviewed publication that we brought out   00:05:52.880 --> 00:05:58.240 at around the time that we were investigating  this initially. And our next steps are   00:05:58.240 --> 00:06:03.840 really to aggregate large data sets across  multiple institutions, have multiple readers   00:06:03.840 --> 00:06:10.640 annotate markup the areas of pneumonia to sort of  solidify the ground truth a little bit, use the CT   00:06:10.640 --> 00:06:17.680 as well and develop this comparable algorithm for  CT in in the process. And ultimately, we want this   00:06:17.680 --> 00:06:22.640 algorithm to be explainable, is really critical  for us to be able to use it clinically is to   00:06:23.760 --> 00:06:28.880 be certain that we're relying on features that  really matter not, not sort of accessory features   00:06:28.880 --> 00:06:34.240 that the neural network sort of coincidentally  saw associated with COVID but actually things   00:06:34.240 --> 00:06:40.320 are related to it and then. And then assess its  clinical utility both in the detection of disease,   00:06:41.120 --> 00:06:45.840 distinguishing between other diseases that are  quite similar like pulmonary edema, and then give   00:06:45.840 --> 00:06:51.360 us best management practices for these patients,  so that's kind of where we're going. Thank you.