WEBVTT 1 00:00:07.466 --> 00:00:08.166 Today, we're going to 2 00:00:08.166 --> 00:00:12.000 be talking about the use of artificial intelligence, and specifically 3 00:00:12.000 --> 00:00:15.533 a convolutional neural network for the detection of intestinal parasites. 4 00:00:15.866 --> 00:00:19.833 And this is, technology that we clinically validated here at ARUP. 5 00:00:19.833 --> 00:00:21.100 We were the first country. 6 00:00:21.100 --> 00:00:23.633 I'm sorry. We're the first lab in the world to do this. 7 00:00:23.633 --> 00:00:27.633 Although now it's actually also been used by the Mayo Clinic there in Rochester. 8 00:00:27.633 --> 00:00:31.933 So they were the second lab to use the technology that we, that we validated. 9 00:00:33.666 --> 00:00:36.666 So here's my disclosures. 10 00:00:36.666 --> 00:00:38.400 Couple of disclosures slide. 11 00:00:38.400 --> 00:00:39.566 So we've got three. 12 00:00:39.566 --> 00:00:41.866 We're going to break this talk into three main sections. 13 00:00:41.866 --> 00:00:44.400 The first half is going to be kind of a review. 14 00:00:44.400 --> 00:00:45.433 It's we're going to recall 15 00:00:45.433 --> 00:00:49.066 the detection of intestinal parasites with an emphasis on microscopy. 16 00:00:49.600 --> 00:00:50.700 And then we're going to describe 17 00:00:50.700 --> 00:00:54.566 the development of machine learning software to detect intestinal parasites. 18 00:00:54.600 --> 00:00:57.533 And my role in this project has been a microbiologist. 19 00:00:57.533 --> 00:01:00.933 So this is going to be very skimming surface stuff. 20 00:01:00.933 --> 00:01:05.200 I'm not going to go into the nuances of model design and PR curves 21 00:01:05.200 --> 00:01:06.833 and all that kind of stuff because, 22 00:01:06.833 --> 00:01:09.600 that's not actually my expertise on the subject, 23 00:01:09.600 --> 00:01:11.966 but I do want to talk to you guys about some of the things 24 00:01:11.966 --> 00:01:14.700 that we had to develop and overcome along the way. 25 00:01:14.700 --> 00:01:17.700 And last, we're going to talk about implementing this technology 26 00:01:17.700 --> 00:01:19.066 in the clinical lab. 27 00:01:19.066 --> 00:01:21.800 The benefits derive there from and maybe some of the challenges 28 00:01:21.800 --> 00:01:24.800 you might have to overcome if you decide you want to do this. 29 00:01:25.366 --> 00:01:26.466 So let's get started. 30 00:01:26.466 --> 00:01:28.500 For first of all, let's go ahead and review the micro, 31 00:01:29.500 --> 00:01:32.566 the micro microscopic diagnostic methods for parasites. 32 00:01:32.566 --> 00:01:36.200 So conventional diagnosis of intestinal parasites. 33 00:01:36.233 --> 00:01:40.366 Well of course the old gold standard for the most part still, 34 00:01:41.566 --> 00:01:44.366 is morphologic analysis and microscopy. 35 00:01:44.366 --> 00:01:48.033 And we'll go over specific methods on, on, upcoming slide. 36 00:01:48.800 --> 00:01:52.533 Later antigen detection assays were developed and these are, 37 00:01:52.666 --> 00:01:53.333 these are good 38 00:01:53.333 --> 00:01:56.866 for some of the high target pathogens things like giardia and crypto. 39 00:01:57.133 --> 00:02:00.933 They're they're often more sensitive than morpho morphology in microscopy. 40 00:02:00.933 --> 00:02:03.966 And they're not quite as expensive as molecular. 41 00:02:04.200 --> 00:02:07.166 They're good if you suspect things like crypto and guardian and immuno 42 00:02:07.166 --> 00:02:10.700 competent patients in the US with no relevant travel history. 43 00:02:10.933 --> 00:02:13.400 Those are going to be your most likely candidates. 44 00:02:13.400 --> 00:02:15.266 And so, 45 00:02:15.266 --> 00:02:18.400 antigen detection is it's good for screening those kind of 46 00:02:18.400 --> 00:02:22.466 suspect organisms before going straight to like an O and P exam, for example. 47 00:02:24.066 --> 00:02:26.133 And last of course is molecular detection. 48 00:02:26.133 --> 00:02:30.100 And molecular detection is again molecular 49 00:02:30.100 --> 00:02:34.033 detection is is now great for a specific parasite targeting. 50 00:02:34.033 --> 00:02:36.966 Some of the broad diarrheal panels have common parasites. 51 00:02:36.966 --> 00:02:39.766 And that's good if you don't have a suspect agent. 52 00:02:39.766 --> 00:02:43.566 But if you actually specifically with suspecting a specific parasite 53 00:02:44.066 --> 00:02:47.533 molecular detection is is still often cost prohibited 54 00:02:47.833 --> 00:02:50.833 compared to, say, antigen detection or microscopy. 55 00:02:51.700 --> 00:02:55.366 And then last, of course, is artificial intelligence and machine learning, 56 00:02:55.766 --> 00:02:59.300 where if you think about it, is really just an updated version 57 00:02:59.300 --> 00:03:02.300 of good old morphology in microscopy. 58 00:03:02.666 --> 00:03:04.500 So what are the some of the morphological 59 00:03:04.500 --> 00:03:07.500 microscopic methods for the detection of parasites? 60 00:03:08.100 --> 00:03:11.733 Well, you have the wet mouthed either director concentrated in the trichomes 61 00:03:11.733 --> 00:03:12.766 smear. 62 00:03:12.766 --> 00:03:16.266 And these two combined form the classic Auvergne parasite exam 63 00:03:16.266 --> 00:03:21.300 for protozoan cysts, protozoan trophies sites, worm eggs and worm larvae. 64 00:03:21.666 --> 00:03:24.633 Then you have your modified acid force and saffron and stains. 65 00:03:24.633 --> 00:03:28.200 These are great for things like the cock citizens like Dois Aspera, 66 00:03:28.500 --> 00:03:31.666 Cyclospora and the Gregor and Cryptosporidium. 67 00:03:32.200 --> 00:03:35.600 Histopathology is used for the detection and identification 68 00:03:35.600 --> 00:03:38.600 of parasites and biopsy specimens. 69 00:03:38.733 --> 00:03:41.033 And then you have the gross examination of helminths. 70 00:03:41.033 --> 00:03:44.166 For example, an adult asterisk gets passed, 71 00:03:44.166 --> 00:03:47.600 or a pin worm found crawling across a poopy diaper. 72 00:03:49.200 --> 00:03:51.900 We're going to focus on these top three web mount, 73 00:03:51.900 --> 00:03:54.900 tri chrome and modified acid fast because those are the, 74 00:03:55.500 --> 00:03:58.566 those are the the methods that we've either clinically validated by 75 00:03:58.566 --> 00:04:01.566 I or we're currently in the process of doing so. 76 00:04:02.500 --> 00:04:07.400 So let's review the, as my friend and medical director, doctor 77 00:04:07.400 --> 00:04:11.733 Mark couturier, likes to say, the insanity of the open and parasite exam. 78 00:04:12.233 --> 00:04:15.766 So your traditional open parasite exam is fixed stool. 79 00:04:15.766 --> 00:04:20.566 Historically, it was a two volume, a two vial system such as formalin and PVA. 80 00:04:20.933 --> 00:04:25.833 Now there's a lot of, environmentally friendly fixative and single vial systems. 81 00:04:26.700 --> 00:04:29.700 The specimen is concentrated for increased sensitivity. 82 00:04:30.200 --> 00:04:33.266 And here at a rapid, we concentrate both the wet mount in the tri chrome, 83 00:04:33.633 --> 00:04:36.433 the wet mouth, which I already kind of alluded to. 84 00:04:36.433 --> 00:04:37.766 That's mixed. 85 00:04:37.766 --> 00:04:40.366 May or may not be mixed with iodine and, 86 00:04:41.933 --> 00:04:42.300 and then the 87 00:04:42.300 --> 00:04:45.300 tri chrome slide with this is a permanent smear. 88 00:04:46.233 --> 00:04:47.900 That's used for the protozoan troughs. 89 00:04:47.900 --> 00:04:50.900 Insists recommended on poo. 90 00:04:50.900 --> 00:04:54.733 So ideally you should have three unique specimens per patient. 91 00:04:54.733 --> 00:04:57.966 And for some of the harder to detect things such as facile 92 00:04:57.966 --> 00:05:00.966 and strongyloides, you may even want to go 5 to 7. 93 00:05:01.200 --> 00:05:04.533 It's not recommended for patients with hospital onset diarrhea. 94 00:05:04.800 --> 00:05:09.300 Ideally, you should only be ordering open parasite exams under people 95 00:05:09.300 --> 00:05:12.300 with a high pretest probability of having a parasitic disease, 96 00:05:12.300 --> 00:05:16.133 such as immunocompromised patients, patients with persistent 97 00:05:16.166 --> 00:05:20.833 chronic diarrhea, pertinent exposure history, and pertinent travel history. 98 00:05:20.833 --> 00:05:23.833 For example, someone who visited or emigrated from 99 00:05:23.900 --> 00:05:26.600 an endemic area. 100 00:05:26.600 --> 00:05:30.833 I'm not going to go into this next slide too much because every lab has 101 00:05:30.833 --> 00:05:35.200 their own nuances for the parasite exam, but I like to emphasize 102 00:05:35.200 --> 00:05:39.433 how much variability there is in this analysis, because, you know, 103 00:05:40.033 --> 00:05:44.566 you have the ratio of stool to fixative, how much goes in the original specimen. 104 00:05:44.566 --> 00:05:47.566 And then, that that might affect, 105 00:05:48.133 --> 00:05:51.766 how much is makes it through the, the, the processing, the spinning. 106 00:05:52.200 --> 00:05:53.566 And then you've got the smear prep. 107 00:05:53.566 --> 00:05:56.833 Different technologies might make their wet mounts at different thicknesses. 108 00:05:57.166 --> 00:05:59.566 Technologists or technicians, 109 00:05:59.566 --> 00:06:01.933 everyone in the lab might make their tray crumbs a little different. 110 00:06:01.933 --> 00:06:04.600 So there's a lot of variability in this assay, 111 00:06:04.600 --> 00:06:08.133 which could ultimately end up affecting the sensitivity of the assay. 112 00:06:09.600 --> 00:06:10.200 When reading an 113 00:06:10.200 --> 00:06:13.233 open parasite exam before we instituted the I, 114 00:06:13.633 --> 00:06:17.266 a typical run today or up was 30 wet mount 30 specimens. 115 00:06:17.266 --> 00:06:20.433 So 30 wet mounts and their corresponding trichomes, which means 116 00:06:20.433 --> 00:06:23.466 the technologist looked at 60 slides per run, 117 00:06:24.066 --> 00:06:27.866 an average reader that would take two and a half to three hours. 118 00:06:27.866 --> 00:06:31.333 And unless you're in a high volume lab, you could have anywhere 119 00:06:31.333 --> 00:06:34.333 from 95 to 98% negative rate. 120 00:06:34.400 --> 00:06:34.666 All right. 121 00:06:34.666 --> 00:06:36.200 Our lab positives are back. 122 00:06:36.200 --> 00:06:38.866 Read every positive is it's back read. 123 00:06:38.866 --> 00:06:41.966 And if there's a discrepancy it's read by either myself or the medical director. 124 00:06:41.966 --> 00:06:43.733 That just adds more time. 125 00:06:43.733 --> 00:06:48.166 So you can see that this is a rather time consuming assay for a rather low yield. 126 00:06:48.600 --> 00:06:49.866 And so, 127 00:06:49.866 --> 00:06:52.933 you know, a technologist could spend anywhere from 2 to 5 minutes a slide 128 00:06:52.933 --> 00:06:55.600 if you have a questionable negative like you thought you saw something, 129 00:06:55.600 --> 00:06:59.700 but you weren't sure and you scrutinize it longer, now you're spending more time. 130 00:06:59.966 --> 00:07:03.633 So the kind of the take home message is a lot of time for little yield. 131 00:07:04.700 --> 00:07:07.666 And there's some other possible concerns for parasite reading, 132 00:07:07.666 --> 00:07:10.966 such as eye strain, neuromuscular strain from, 133 00:07:12.333 --> 00:07:16.200 adjusting the microscope, burnout and satisfaction. 134 00:07:16.533 --> 00:07:18.700 Like, do you really want to sit at a, at a, 135 00:07:18.700 --> 00:07:22.333 at a microscope eight hours a day for just looking at negative poop? 136 00:07:22.333 --> 00:07:24.700 Maybe for your occasional blastocyst? 137 00:07:24.700 --> 00:07:26.066 This or giardia. 138 00:07:26.066 --> 00:07:27.033 And then there's a lot of things 139 00:07:27.033 --> 00:07:31.066 that might affect accuracy, such as experience, rest, distractions. 140 00:07:31.066 --> 00:07:32.833 Are you constantly having to answer the phone? 141 00:07:32.833 --> 00:07:33.566 Are your colleagues 142 00:07:33.566 --> 00:07:37.300 constantly coming up asking you questions while you're trying to read this run? 143 00:07:37.800 --> 00:07:39.000 Am versus PM? 144 00:07:39.000 --> 00:07:41.100 Me personally, I prefer to do my own 145 00:07:41.100 --> 00:07:44.100 reading in the morning because I know in the afternoon I get more tired, 146 00:07:44.200 --> 00:07:47.200 I get more eyestrain, and I'm probably less sensitive. 147 00:07:47.900 --> 00:07:49.500 Multiple runs throughout the day. 148 00:07:49.500 --> 00:07:51.300 The more runs you do in the day, the likely 149 00:07:51.300 --> 00:07:54.433 your sensitivity is going to go down and later runs. 150 00:07:55.066 --> 00:07:58.900 And then you have things like low parasite burden challenges, interpretation. 151 00:07:59.300 --> 00:08:02.400 And what I mean by that is if there is a bias 152 00:08:02.400 --> 00:08:05.400 or there is a common error made in a lab, something, 153 00:08:06.200 --> 00:08:09.900 a common dogmatic mistake that's made that gets perpetuated over time, 154 00:08:10.166 --> 00:08:12.566 it can talk to new lab staff, and next thing you know, 155 00:08:12.566 --> 00:08:15.733 whatever this error is, it's now become culture in your lab. 156 00:08:17.366 --> 00:08:20.100 So how can we make this process, 157 00:08:20.100 --> 00:08:24.000 more efficient, more accurate, and possibly more fun? 158 00:08:24.000 --> 00:08:27.833 Well, that's where the digital imaging and machine learning comes into play. 159 00:08:28.600 --> 00:08:31.200 Now I'm going to use some terminology in this talk. 160 00:08:31.200 --> 00:08:33.833 So I want to define a little bit of it upfront. 161 00:08:33.833 --> 00:08:36.733 In case I forget to describe it detail later. 162 00:08:36.733 --> 00:08:39.733 CNN is a convolutional neural network. 163 00:08:40.000 --> 00:08:42.400 A class is 164 00:08:42.400 --> 00:08:44.500 it's a classification based on shared features. 165 00:08:44.500 --> 00:08:46.933 So when you're training software, you create classes. 166 00:08:46.933 --> 00:08:49.633 So a G assist is a class. 167 00:08:49.633 --> 00:08:51.100 And ask Iris a actually. 168 00:08:51.100 --> 00:08:54.466 And ask or say could be multiple classes because the fertile and infertile 169 00:08:54.466 --> 00:08:57.100 looks so different. Those should actually be two classes. 170 00:08:57.100 --> 00:08:59.700 The end user will see him. It was one organism. 171 00:08:59.700 --> 00:09:01.833 But when you're training you want to classify 172 00:09:01.833 --> 00:09:04.133 things based on shared morphologic features. 173 00:09:05.266 --> 00:09:06.366 Class confusion. 174 00:09:06.366 --> 00:09:07.233 You see this? 175 00:09:07.233 --> 00:09:10.800 This is when the software might flag an organism, but mis identifies it. 176 00:09:11.233 --> 00:09:15.000 There's a lot of, things that can happen that can cause this. 177 00:09:15.000 --> 00:09:17.666 One of the most common is a heavy parasite load. 178 00:09:17.666 --> 00:09:20.833 If the slide is loaded with parasites, the software finds, 179 00:09:20.966 --> 00:09:23.966 but it's almost like it doesn't know what to do because it gets overloaded. 180 00:09:24.100 --> 00:09:26.400 And so it starts putting them in different classes. 181 00:09:26.400 --> 00:09:29.600 But there's ways to overcome all this, and we'll discuss those as we go. 182 00:09:30.066 --> 00:09:31.866 Failed scan an incomplete scan. 183 00:09:31.866 --> 00:09:33.533 That's when your scanner 184 00:09:33.533 --> 00:09:36.900 either can't scan the slide or can't scan the entire thing. 185 00:09:36.933 --> 00:09:40.100 That's usually associated with an insufficient 186 00:09:40.100 --> 00:09:43.100 amount of fecal material on the slide. 187 00:09:43.266 --> 00:09:44.500 So digital imaging. 188 00:09:44.500 --> 00:09:46.433 So in order for this 189 00:09:46.433 --> 00:09:49.600 to really be effective in the lab it's got to meet certain criteria. 190 00:09:49.966 --> 00:09:52.800 So first of all must be high enough resolution for fine 191 00:09:52.800 --> 00:09:54.600 detailed determination. 192 00:09:54.600 --> 00:09:56.700 You need to see that in to me bassist. 193 00:09:56.700 --> 00:09:58.533 The intermediate nucleus. You need to see the 194 00:09:59.533 --> 00:10:00.866 maybe the fibrils are the media. 195 00:10:00.866 --> 00:10:04.366 You need to see something that you can look at that image and say, yeah, 196 00:10:04.766 --> 00:10:07.400 that's the suspect. Parasite. 197 00:10:07.400 --> 00:10:10.866 The must improve ease of review, and it can't be too cumbersome. 198 00:10:11.400 --> 00:10:14.700 The, the scan time must be time efficient. 199 00:10:14.700 --> 00:10:17.700 And I'll go over this on a later slide in more detail. 200 00:10:18.166 --> 00:10:19.166 It must be user friendly. 201 00:10:19.166 --> 00:10:20.233 Especially the end result. 202 00:10:20.233 --> 00:10:22.400 When the technologist is analyzing the images. 203 00:10:22.400 --> 00:10:25.066 It must be very user friendly and streamlined. 204 00:10:25.066 --> 00:10:26.233 If it becomes too cumbersome 205 00:10:26.233 --> 00:10:29.700 and burdensome, it's not going to help with the turnaround time issues. 206 00:10:30.166 --> 00:10:33.633 And ideally, the images should be equal or better than what's seen through an IPS. 207 00:10:33.633 --> 00:10:37.100 Now they're not going to be as good as what you might find in our 208 00:10:37.700 --> 00:10:42.266 textbook or the DPG website, but they can be pretty darn 209 00:10:42.266 --> 00:10:46.100 good enough that a competent technologist could look at an image and say, 210 00:10:46.400 --> 00:10:47.966 yeah, I think that's suspicious. 211 00:10:49.300 --> 00:10:52.300 So some concepts going into development. 212 00:10:52.500 --> 00:10:56.233 First of all, everyone needs to understand this is not a definitive parasite 213 00:10:56.233 --> 00:10:57.900 identification tool. 214 00:10:57.900 --> 00:11:00.866 This is a screening tool. 215 00:11:00.866 --> 00:11:01.933 Basically. 216 00:11:01.933 --> 00:11:05.400 And I'll go more into sensitivity and specificity training in a second. 217 00:11:05.800 --> 00:11:10.233 But I the software is going to flag something on every single specimen. 218 00:11:10.533 --> 00:11:14.033 And basically the end user goes through the images and determines 219 00:11:14.033 --> 00:11:18.300 if the slide needs to be manually pulled, that ultimately 220 00:11:18.300 --> 00:11:22.100 the definitive identification comes from the manual back reading of a slide. 221 00:11:22.333 --> 00:11:25.333 If you think of it this way, say you have 20 slides on your run, 222 00:11:25.633 --> 00:11:28.566 so you've got these these scanned, 20 scanned. 223 00:11:28.566 --> 00:11:31.566 Try chromes from from your own pee run. 224 00:11:31.800 --> 00:11:32.866 The software says. Yeah. 225 00:11:32.866 --> 00:11:35.800 So these three over here, these have suspect organism. 226 00:11:35.800 --> 00:11:40.800 Let's pull these for manual read and specific scrutinizing these other 17. 227 00:11:41.166 --> 00:11:44.166 You can report those out as negative without having to read a slide. 228 00:11:44.433 --> 00:11:49.800 The idea was if we can get the software or at least image analysis I should say. 229 00:11:50.100 --> 00:11:52.200 Because another important thing to understand is 230 00:11:52.200 --> 00:11:54.366 the software doesn't determine this is negative. 231 00:11:54.366 --> 00:11:56.500 And report it out through a reporting system. 232 00:11:56.500 --> 00:11:59.100 The technologist still needs to look at all the images. 233 00:11:59.100 --> 00:12:02.800 But the end goal was if you can have 70 to 80% of the clean 234 00:12:02.800 --> 00:12:06.466 negatives ruled out by image analysis, and you didn't have to read the slide, 235 00:12:06.900 --> 00:12:11.066 I can tell you that a clock in our actual practice 236 00:12:11.400 --> 00:12:15.900 for tri and modified acid fast, it's probably closer to 9,095%. 237 00:12:15.900 --> 00:12:19.366 The technologies are most of the thing the technologists are 238 00:12:19.533 --> 00:12:20.533 are pulling the back. 239 00:12:20.533 --> 00:12:22.366 Read our true positives. 240 00:12:23.366 --> 00:12:26.366 And right now in our lab, the web mount is still read normally 241 00:12:26.433 --> 00:12:29.700 because we haven't clinically validated the web mount model yet, 242 00:12:29.900 --> 00:12:33.100 but we're working on it and that's what I'll finish my talk with. 243 00:12:33.600 --> 00:12:37.266 The. The basic thing to remember is this is to compliment the technologist, 244 00:12:37.266 --> 00:12:39.200 not to replace the technologist. 245 00:12:39.200 --> 00:12:42.300 It's not going to replace the need for proficiency or competency, 246 00:12:42.600 --> 00:12:45.633 because you're going to still need a competent parasitologist to analyze 247 00:12:45.633 --> 00:12:49.433 the images, and a competent parasitologist to back read the positive. 248 00:12:49.633 --> 00:12:52.466 So this is not going to negate the importance of having 249 00:12:52.466 --> 00:12:55.466 trained parasitologist in your lab. 250 00:12:55.700 --> 00:12:58.700 So I'm going to go over this slide really quick. 251 00:13:00.066 --> 00:13:03.366 To kind of basically described how we train the software. 252 00:13:03.700 --> 00:13:07.333 We used what's called supervised machine learning. 253 00:13:07.333 --> 00:13:10.933 And so what that is is you start with a slide. 254 00:13:11.766 --> 00:13:14.466 Maybe it has gorilla IoT amoeba. 255 00:13:14.466 --> 00:13:16.766 And to me it has a parasite on it. 256 00:13:16.766 --> 00:13:21.533 And you scan that, you let the software scan it and then you go into the raw scan 257 00:13:21.933 --> 00:13:25.500 and the technologist boxes say it's a job, a slide. 258 00:13:25.500 --> 00:13:28.500 You box 100 examples of giardia, 259 00:13:28.533 --> 00:13:32.066 and then so you're telling the software, this is what Giardia looks like. 260 00:13:32.700 --> 00:13:36.433 Then you have the, the scanner re scan that slide. 261 00:13:36.433 --> 00:13:39.333 Or actually you could even go back to the original scan 262 00:13:39.333 --> 00:13:42.333 and you tell the software, okay, this is what Giardia looks like. 263 00:13:42.600 --> 00:13:44.133 Find me more examples. 264 00:13:44.133 --> 00:13:48.633 And then the software comes back with like 500 hits that it thinks is great 265 00:13:48.666 --> 00:13:52.866 based on what you said giardia should look like the technologist 266 00:13:52.866 --> 00:13:55.866 or the the person trading the software 267 00:13:55.900 --> 00:13:59.900 then goes in and looks at all those images that the software thinks is great, 268 00:14:00.600 --> 00:14:04.866 and it says, and if it if it's a good idea and you think it'd be good for training, 269 00:14:04.866 --> 00:14:06.333 you confirm it. 270 00:14:06.333 --> 00:14:09.433 If it looks like it's diarrhea, but it's kind of crappy, and you 271 00:14:09.600 --> 00:14:11.833 you don't think it'd be a good training exemplar? 272 00:14:11.833 --> 00:14:15.266 We would classify it as unknown and that hides it from the software. 273 00:14:15.266 --> 00:14:17.533 So they don't so it doesn't see it again. 274 00:14:17.533 --> 00:14:21.566 And if it's a piece of junk, you could either call it bad box or junk, 275 00:14:21.800 --> 00:14:24.966 or if you find the same artifact is constantly being classified 276 00:14:24.966 --> 00:14:29.100 as a parasite, you can give it a formal name like Blobby Blue 277 00:14:29.100 --> 00:14:33.000 Object one, and then you hide it in the background as an antique class. 278 00:14:33.000 --> 00:14:36.000 So the software classifies it and groups it, 279 00:14:36.066 --> 00:14:37.933 but it doesn't show it to the end user. 280 00:14:38.933 --> 00:14:42.400 And basically you keep doing this over and over again in the software. 281 00:14:42.400 --> 00:14:45.400 It gets better and better at identifying the, 282 00:14:45.666 --> 00:14:48.766 of, of of, of, of detecting the parasite. 283 00:14:48.766 --> 00:14:51.200 I should just say detect, not identify. 284 00:14:51.200 --> 00:14:53.566 And you use holdout slides to test this. 285 00:14:53.566 --> 00:14:57.300 So you have some slides set aside specifically for analyzing the model. 286 00:14:57.600 --> 00:15:00.333 And every time you add more data or you update the model, you run this, 287 00:15:00.333 --> 00:15:03.666 this holdout, slide through and see if it performs equally. 288 00:15:03.900 --> 00:15:06.833 If it starts to find more organisms, then you could say that 289 00:15:06.833 --> 00:15:08.633 your model is actually getting better. 290 00:15:08.633 --> 00:15:11.033 But say it all of a sudden. Slides a thousand things. 291 00:15:11.033 --> 00:15:11.866 That's all junk. 292 00:15:11.866 --> 00:15:14.100 Well, something happened goofy with your model, 293 00:15:14.100 --> 00:15:16.600 and then you have to go back and troubleshoot it. 294 00:15:16.600 --> 00:15:18.400 Eventually you start adding, 295 00:15:20.200 --> 00:15:21.400 additional slides 296 00:15:21.400 --> 00:15:24.533 for each species, maybe different stains, different fixative. 297 00:15:24.533 --> 00:15:26.400 So you get some variation. 298 00:15:26.400 --> 00:15:29.400 And then the model can learn to recognize in this case intermedia, 299 00:15:29.600 --> 00:15:32.600 it doesn't matter what it looks like, it can call all of these into me. 300 00:15:33.100 --> 00:15:34.833 But maybe it flagged something as an to me. 301 00:15:34.833 --> 00:15:35.800 But that's not an to me. 302 00:15:35.800 --> 00:15:38.666 But in this case, it's kind of a mouse that you're like, well, you know what? 303 00:15:38.666 --> 00:15:40.300 We want to train on a mouse sticks. 304 00:15:40.300 --> 00:15:43.300 So we're going to 305 00:15:43.866 --> 00:15:44.833 there we go. Sorry. 306 00:15:44.833 --> 00:15:46.900 We're going to add that back to our data matrix. 307 00:15:46.900 --> 00:15:48.600 And the process continues. 308 00:15:48.600 --> 00:15:50.200 So that's supervised machine learning. 309 00:15:50.200 --> 00:15:52.500 And you can just kind of do it over and over again. 310 00:15:52.500 --> 00:15:56.000 Ideally you'd need about 500 to 1000 exemplars 311 00:15:56.233 --> 00:15:59.233 before your training before the software gets really good. 312 00:15:59.366 --> 00:16:02.400 You could also classify things on the confidence level. 313 00:16:02.733 --> 00:16:04.233 And so you could tell the software, hey, 314 00:16:04.233 --> 00:16:07.066 don't show me anything but the low 90% confidence. 315 00:16:07.066 --> 00:16:10.766 And that increases the chances of what's going to be flagged as legitimate. 316 00:16:12.000 --> 00:16:14.200 So when we first started this, 317 00:16:14.200 --> 00:16:17.766 these were the classes we wanted to train because keeping in mind 318 00:16:17.766 --> 00:16:20.900 that we still read the web now, normally we weren't as concerned 319 00:16:20.900 --> 00:16:24.266 with cysts, so we focused mostly on trophies. 320 00:16:24.266 --> 00:16:26.933 Oh, it's especially the high yield pathogens. 321 00:16:26.933 --> 00:16:29.633 So this are the classes that were trained. 322 00:16:29.633 --> 00:16:31.200 And this is what the end user see. 323 00:16:31.200 --> 00:16:34.300 So even though assistant troughs were trained separately, 324 00:16:34.766 --> 00:16:37.766 the end user just seized because at the end of the day 325 00:16:38.366 --> 00:16:41.366 what what's most important is does a patient of GRA, 326 00:16:41.700 --> 00:16:44.966 we also trained on red cells and white cells because our lab reports 327 00:16:44.966 --> 00:16:49.633 those at the three and four plus level yeasts were trained as an antique class. 328 00:16:49.633 --> 00:16:51.933 So our lab doesn't report yeast, 329 00:16:51.933 --> 00:16:55.200 but the software will flag the yeast, but it hides it in the background. 330 00:16:55.533 --> 00:16:58.766 We had to do that because if you had a yeast with bacteria 331 00:16:58.766 --> 00:17:02.966 ringing around it, it kept class confusing that for blastocyst, this 332 00:17:04.733 --> 00:17:07.500 and this is what we did for the modified acid fast. 333 00:17:07.500 --> 00:17:09.900 We trained on Cyclospora and Cryptosporidium. 334 00:17:09.900 --> 00:17:11.900 Only we didn't train on Cisco. 335 00:17:11.900 --> 00:17:16.433 I saw espera because in our lab, every time a specimen comes in 336 00:17:16.433 --> 00:17:20.633 for the modified acid fast, we also do a wet mount UV screen, 337 00:17:20.633 --> 00:17:24.533 which increases sensitivity in a system I saw Spurs president. 338 00:17:24.566 --> 00:17:26.500 It's going to get picked up in the oven. 339 00:17:26.500 --> 00:17:29.800 I don't know if you guys know or not, but UV is is highly sensitive 340 00:17:29.800 --> 00:17:34.100 for screening things like cyclospora and Cisco I saw espera. 341 00:17:34.333 --> 00:17:36.333 It doesn't work for cryptosporidium, 342 00:17:36.333 --> 00:17:39.800 but it's much more sensitive than the modified acid test with the San Fernan. 343 00:17:40.266 --> 00:17:41.266 Interestingly enough, 344 00:17:41.266 --> 00:17:44.700 in the last couple of years, we've gotten about 8 to 10 specimens of cystitis. 345 00:17:44.700 --> 00:17:45.533 Aspera. 346 00:17:45.533 --> 00:17:48.666 So we're thinking about redoing this model and adding it back in. 347 00:17:50.600 --> 00:17:52.500 So with regards to sensitivity 348 00:17:52.500 --> 00:17:57.266 and specificity, how how great do you want the model to be? 349 00:17:57.566 --> 00:18:01.500 Well, if you train only on really high quality 350 00:18:01.500 --> 00:18:04.500 super textbook examples. 351 00:18:05.033 --> 00:18:06.133 Well, the good news is, 352 00:18:06.133 --> 00:18:09.400 is everything that the software will flag is likely to be true. 353 00:18:09.900 --> 00:18:14.900 But you're going to miss low positives or poorly preserved, specimens 354 00:18:14.900 --> 00:18:16.266 that might be poorly fixed 355 00:18:16.266 --> 00:18:19.266 because maybe there was a delay getting the stool into the fixative. 356 00:18:19.666 --> 00:18:22.133 So ideally 357 00:18:22.133 --> 00:18:25.933 you want it if you go to, if you go 358 00:18:25.966 --> 00:18:30.066 higher sensitivity or specificity, it starts flagging too much stuff. 359 00:18:30.066 --> 00:18:33.300 And next thing you know, your technologist is spending too much time 360 00:18:33.300 --> 00:18:36.700 going through junk, trying to find the true parasites. 361 00:18:36.933 --> 00:18:39.866 So you kind of want to find that sweet spot where 362 00:18:39.866 --> 00:18:43.166 you're sure to get all the positives, but, 363 00:18:44.700 --> 00:18:45.166 it still 364 00:18:45.166 --> 00:18:48.166 flags some junk, but not enough to become overwhelming. 365 00:18:48.266 --> 00:18:49.000 And I'll show you. 366 00:18:49.000 --> 00:18:50.700 I'll show you a screenshot of like, 367 00:18:50.700 --> 00:18:54.166 a clean negative slide so you can kind of under understand how, 368 00:18:54.733 --> 00:18:57.733 how this looks to the technologist. 369 00:18:58.200 --> 00:19:00.333 So designing a scan area. 370 00:19:00.333 --> 00:19:03.766 So when you're deciding how much of the tray 371 00:19:03.766 --> 00:19:06.766 crumb slide you want to scan, you 372 00:19:06.933 --> 00:19:10.266 if you scan the whole thing that's going to take 25 to 30 minutes. 373 00:19:10.266 --> 00:19:11.166 That's not efficient. 374 00:19:11.166 --> 00:19:15.800 When a normal technologist reads it in just a couple minutes, you want something. 375 00:19:15.800 --> 00:19:20.500 Essentially, you want a scan area that, covers 376 00:19:20.500 --> 00:19:24.433 the minimum number of required fields to be read and on a track or on slide. 377 00:19:24.433 --> 00:19:27.533 That's typically 100 fields at a, 378 00:19:28.300 --> 00:19:32.233 100 X with oil, but should still be under five minutes. 379 00:19:32.233 --> 00:19:37.366 Well, we were able to develop a scan area that takes roughly 4 to 5 minutes. 380 00:19:37.366 --> 00:19:40.000 And it's about I think it's about I don't remember the numbers 381 00:19:40.000 --> 00:19:44.100 because it's been a while, but it's 11mm by 2 or 3mm. 382 00:19:44.466 --> 00:19:48.033 So it covers way more than the minimum read requirements 383 00:19:48.033 --> 00:19:52.633 for a technologist, but still does it about the same amount of time. 384 00:19:53.700 --> 00:19:54.266 And we also 385 00:19:54.266 --> 00:19:58.233 prepare our smears with the fecal material focused in one area. 386 00:19:58.233 --> 00:20:01.233 So the scan area could be set. 387 00:20:01.566 --> 00:20:03.833 We did some limit of detection studies 388 00:20:03.833 --> 00:20:06.833 to to, to test the scan area. 389 00:20:07.000 --> 00:20:08.400 And this is one run. 390 00:20:08.400 --> 00:20:09.800 But we did it five times. 391 00:20:09.800 --> 00:20:12.966 Basically I took mixture and Blastocysts. 392 00:20:13.333 --> 00:20:14.833 I created fecal sessions 393 00:20:14.833 --> 00:20:18.700 and they were intermixed with the daily op workflow in our lab. 394 00:20:19.033 --> 00:20:22.700 And so the technologists read these as if they were normal patient specimens. 395 00:20:22.966 --> 00:20:25.300 We also ran them on the software. 396 00:20:25.300 --> 00:20:26.233 And all this is on. 397 00:20:26.233 --> 00:20:29.633 And although this is only one example, this is pretty much, 398 00:20:30.833 --> 00:20:32.666 par for what we saw. 399 00:20:32.666 --> 00:20:36.900 The technologist rarely detected anything below 1 to 16, 400 00:20:37.233 --> 00:20:42.100 but the software typically found at least one organism all the way up to 1 to 256. 401 00:20:42.400 --> 00:20:43.966 And we did this five times through. 402 00:20:43.966 --> 00:20:45.366 So we were pretty confident, 403 00:20:46.933 --> 00:20:49.333 that it was more sensitive than the read. 404 00:20:49.333 --> 00:20:52.633 The modified acid fast was interesting. 405 00:20:53.600 --> 00:20:55.200 It was similar results. 406 00:20:55.200 --> 00:20:58.366 But one caveat when we did this, 407 00:20:58.666 --> 00:21:01.933 the technologist only had the modified acid fast to go by. 408 00:21:02.233 --> 00:21:03.866 They didn't have the UV screen. 409 00:21:03.866 --> 00:21:06.066 If this was a normal clinical patient, 410 00:21:06.066 --> 00:21:09.066 they probably would have found a lot more of these by UV screen. 411 00:21:09.100 --> 00:21:14.666 But we were comparing modified acid test technologies, modified acid fast software. 412 00:21:14.933 --> 00:21:17.933 So we weren't really trying to, 413 00:21:19.033 --> 00:21:21.200 we didn't want to bring UV into the equation. 414 00:21:21.200 --> 00:21:24.066 So we kind of had to do this one a little bit different. 415 00:21:24.066 --> 00:21:28.000 But needless to say, the software is more sensitive 416 00:21:28.000 --> 00:21:31.000 than the technologist. 417 00:21:31.433 --> 00:21:34.766 So we talked about we reviewed the, the, the 418 00:21:35.100 --> 00:21:38.166 we reviewed the, the, kind of the historical overview 419 00:21:38.166 --> 00:21:41.633 of morphologic diagnosis of intestinal parasites. 420 00:21:42.100 --> 00:21:43.866 I went over a little quickly, 421 00:21:46.500 --> 00:21:47.466 challenges with 422 00:21:47.466 --> 00:21:51.400 designing a model for the detection of these parasites. 423 00:21:51.600 --> 00:21:55.566 Now we're going to now we're going to talk about actually applying this in the lab. 424 00:21:55.900 --> 00:21:59.500 So I'll talk about some, some, challenges to consider 425 00:21:59.500 --> 00:22:01.633 and what you might have to do to overcome them. 426 00:22:01.633 --> 00:22:04.733 And then I'm going to walk you through a couple of screenshot examples 427 00:22:04.733 --> 00:22:08.100 of what our technologies actually see when they do this. 428 00:22:08.900 --> 00:22:13.266 So first of all, let's go back to the parasite workflow. 429 00:22:13.733 --> 00:22:17.300 What are some challenges to consider in your workflow process? 430 00:22:18.600 --> 00:22:22.600 Well, let's talk about potential changes to open parasite processing. 431 00:22:22.933 --> 00:22:25.500 First of all is slide preparation. 432 00:22:25.500 --> 00:22:28.400 Some labs like to do the old hills and valleys 433 00:22:28.400 --> 00:22:31.400 where you have varying areas of thickness and sin. 434 00:22:31.633 --> 00:22:34.166 That's not good for the scanner you are. 435 00:22:34.166 --> 00:22:38.066 You want to create a nice homogenous slide smear. 436 00:22:38.300 --> 00:22:41.800 So so, so when the scanner is going 437 00:22:41.800 --> 00:22:45.733 across the slide, everything's at a relatively the same focal play. 438 00:22:45.733 --> 00:22:47.666 Now, stools are very 439 00:22:49.300 --> 00:22:51.300 un uniform matrix. 440 00:22:51.300 --> 00:22:54.933 There's always going to be variability even on this one slide based on 441 00:22:55.133 --> 00:22:56.400 chunks of fecal material. 442 00:22:56.400 --> 00:22:58.200 So it's never going to be perfect. 443 00:22:58.200 --> 00:23:01.600 But ideally you want to make your smear preps as thin as possible. 444 00:23:01.600 --> 00:23:05.733 So if your lab is doing the hills and valleys, your technicians or technologist, 445 00:23:05.733 --> 00:23:09.166 whoever makes these smears is probably is going to have to, 446 00:23:10.600 --> 00:23:13.466 develop, 447 00:23:13.466 --> 00:23:15.733 adopt a new technique. 448 00:23:15.733 --> 00:23:18.300 Cover slipping is required for scanning. 449 00:23:18.300 --> 00:23:20.900 So you're going to have to cover slip your slides. 450 00:23:20.900 --> 00:23:23.900 We use an automated cover slipper, 451 00:23:24.266 --> 00:23:26.766 that's designed for histopathology. 452 00:23:26.766 --> 00:23:30.733 I would not recommend using things like per mountain glass cover slip slides 453 00:23:31.200 --> 00:23:34.266 because the the warmth of the scanner 454 00:23:34.266 --> 00:23:37.900 or the light can cause that permanent to soften. 455 00:23:38.333 --> 00:23:41.700 And unless your slide is really, really dry and with per melt, 456 00:23:41.700 --> 00:23:46.300 that could really take 2 or 3 days to be like dry enough for a scanner. 457 00:23:46.866 --> 00:23:49.566 Then you might get the, the, 458 00:23:49.566 --> 00:23:53.400 the mounting media might like contaminate and gum up your scanner. 459 00:23:53.400 --> 00:23:56.400 And we actually discovered that the hard way during development. 460 00:23:56.400 --> 00:23:57.833 So you're probably going 461 00:23:57.833 --> 00:24:01.633 to have to invest in or borrow from another lab in your institute. 462 00:24:01.866 --> 00:24:03.366 An automated cover slipper. 463 00:24:05.566 --> 00:24:05.866 You might 464 00:24:05.866 --> 00:24:09.300 want to organize your workflow, your workforce. 465 00:24:09.600 --> 00:24:10.733 How many technologies? 466 00:24:10.733 --> 00:24:13.666 Our lab in our lab, because we're a large reference lab. 467 00:24:13.666 --> 00:24:16.766 Technicians do a lot of prep and technologists do the reading. 468 00:24:17.166 --> 00:24:18.166 But like, as we're seeing 469 00:24:18.166 --> 00:24:21.633 with the wet mount develop, it's going to be a lot more technician heavy. 470 00:24:22.133 --> 00:24:25.433 And so maybe as a technologist retires you might want to replace them 471 00:24:25.433 --> 00:24:26.800 with the technician. 472 00:24:26.800 --> 00:24:30.300 So your if the balance of employees matches 473 00:24:30.700 --> 00:24:33.466 well whatever you need for the workflow in this process, it's 474 00:24:33.466 --> 00:24:38.666 going to vary based on mostly based on your how many specimens you have. 475 00:24:38.766 --> 00:24:41.566 The more specimens you have, 476 00:24:41.566 --> 00:24:45.000 the larger your volume, the more likely you might find, 477 00:24:45.900 --> 00:24:48.900 you need a heavier workforce on the processing in, 478 00:24:50.500 --> 00:24:53.233 developing and maintaining QC for the scans. 479 00:24:53.233 --> 00:24:56.033 Now, this is not the same as QC for the stain. 480 00:24:56.033 --> 00:24:57.233 When you make your trigger 481 00:24:57.233 --> 00:25:01.200 stain every day, you do a daily QC to test for the stain. 482 00:25:01.366 --> 00:25:04.366 But that slide is not used for QC in the scan. 483 00:25:04.466 --> 00:25:09.133 We actually scan a slide every day or every day that scans are done. 484 00:25:09.566 --> 00:25:14.366 And then, whoever's assigned to a particular bench will go in and analyze. 485 00:25:14.666 --> 00:25:17.866 They'll look at how many of the given organisms were flagged. 486 00:25:17.866 --> 00:25:20.866 In the case of try Chrome or QC slide to do a slide. 487 00:25:21.033 --> 00:25:25.366 So like say it's supposed to find 5 to 700 cysts every day. 488 00:25:25.766 --> 00:25:29.566 If it's, you know, 607 50 okay that's fine. 489 00:25:29.800 --> 00:25:32.700 But if it starts finding more or less you might 490 00:25:32.700 --> 00:25:36.000 something might be going on wrong, either with the scanner or the model. 491 00:25:36.433 --> 00:25:37.133 It's funny. 492 00:25:37.133 --> 00:25:38.200 Funny story behind this. 493 00:25:38.200 --> 00:25:42.600 On our modified acid fast, it started finding tons and tons of our crypto 494 00:25:42.600 --> 00:25:43.800 like in the thousands. 495 00:25:43.800 --> 00:25:46.233 And we're like, is the model learning? 496 00:25:46.233 --> 00:25:49.700 It's like, no, the model is locked down before it goes into clinical practice. 497 00:25:49.966 --> 00:25:52.366 Why on earth is it finding more crypto? 498 00:25:52.366 --> 00:25:56.933 Well, the problem is, is every time you scan that slide, it exposes it to light. 499 00:25:56.933 --> 00:26:00.366 And so the light was fading the fecal material in the background, 500 00:26:00.700 --> 00:26:04.900 and it was exposing osis that were previously hidden by fecal material. 501 00:26:05.400 --> 00:26:08.566 So we recommended putting a, I recommended 502 00:26:08.566 --> 00:26:12.300 putting a one month expiration on a QC slide. 503 00:26:12.500 --> 00:26:15.500 And what I'd recommend doing is if you want to do this is 504 00:26:15.800 --> 00:26:18.700 if you have a good specimen, make like 100 slides. 505 00:26:18.700 --> 00:26:20.400 So they all have the same lot. 506 00:26:20.400 --> 00:26:23.266 And just every month replace that slide 507 00:26:23.266 --> 00:26:26.533 that way it'll be a long time before you have to go and prep more slides. 508 00:26:27.933 --> 00:26:30.300 How to handle failed or incomplete scans. 509 00:26:30.300 --> 00:26:31.800 Failed scans don't get scanned. 510 00:26:31.800 --> 00:26:35.200 So in our lab, they go, they go right for manual read. 511 00:26:35.200 --> 00:26:38.500 They just have to be read per convention, an incomplete scan. 512 00:26:38.500 --> 00:26:42.333 To me, an incomplete scan is anything where more than 20% of the slide 513 00:26:42.333 --> 00:26:43.700 does it scan. 514 00:26:43.700 --> 00:26:47.300 So whenever I was doing reads in the lab, because sometimes I still go back 515 00:26:47.300 --> 00:26:50.300 and help the clinical lab, I would actually, 516 00:26:52.433 --> 00:26:55.166 I would, I would look at the whole slide scan 517 00:26:55.166 --> 00:26:58.400 and if it looked like more than 20% of it didn't scanned, 518 00:26:58.400 --> 00:27:01.433 I would just manually read the slide just to be safe 519 00:27:02.600 --> 00:27:04.100 and have a backup plan. 520 00:27:04.100 --> 00:27:06.066 If there's any problems in any step, 521 00:27:06.066 --> 00:27:09.766 like your cover slipper goes down, your scanners broken, 522 00:27:10.100 --> 00:27:13.633 the internet's down so you can't access the software analysis. 523 00:27:13.900 --> 00:27:17.000 Luckily, your specimens are still processed 524 00:27:17.000 --> 00:27:20.133 through normal of parasite procedures, so worst case 525 00:27:20.133 --> 00:27:23.400 scenario, you just go back to manually reading for a couple days. 526 00:27:24.200 --> 00:27:28.000 And another thing to consider is how well will your employees embrace this change? 527 00:27:28.400 --> 00:27:31.700 Like maybe your older workforce, like younger, the younger, 528 00:27:32.233 --> 00:27:35.266 the younger technologists are a little more tech savvy. 529 00:27:35.266 --> 00:27:39.300 They might be more likely to adopt this technology than an older workforce. 530 00:27:39.300 --> 00:27:41.300 Although we found in our lab 531 00:27:41.300 --> 00:27:45.233 it was universally loved across the border, across the board, 532 00:27:45.233 --> 00:27:49.233 whether you were doing on PS for six months or 30 years. 533 00:27:50.866 --> 00:27:52.766 So what's the normal right now? 534 00:27:52.766 --> 00:27:55.066 This is our our lab workflow. 535 00:27:55.066 --> 00:27:57.900 We read the wet mounts per normal 536 00:27:57.900 --> 00:28:01.533 and then you analyze the images and I'll show you some slides 537 00:28:01.533 --> 00:28:04.900 kind of going through the steps of image analysis in just a second. 538 00:28:05.566 --> 00:28:09.633 And you would manually back read the try Chrome or the modified acid fast. 539 00:28:09.633 --> 00:28:13.200 If suspect organisms were seen in the images. 540 00:28:13.800 --> 00:28:16.900 And if there something where you see it in the images, 541 00:28:16.900 --> 00:28:18.333 but you can't find it on the slide. 542 00:28:18.333 --> 00:28:20.433 And I'll show you a neat example of that. 543 00:28:20.433 --> 00:28:23.633 The medical director can analyze the images and make the final call. 544 00:28:24.266 --> 00:28:28.100 Or if there's discrepant results between the wet melt in the images, say you swear 545 00:28:28.100 --> 00:28:31.800 you saw assists on the wet mount, but the software didn't find them. 546 00:28:32.100 --> 00:28:33.300 I would manually read that. 547 00:28:33.300 --> 00:28:36.533 Try Chrome just to make sure that the objects in the wet mount 548 00:28:36.533 --> 00:28:41.566 word artifact, or if you have failed, incomplete or invalid scans. 549 00:28:42.133 --> 00:28:46.833 Again, the end goals were successfully detect common intestinal protozoa 550 00:28:47.333 --> 00:28:51.766 and try to screen out as many as 70 to 80% of the negatives. 551 00:28:52.033 --> 00:28:55.266 And as I said in our lab, it was probably closer to 90 to 95. 552 00:28:56.500 --> 00:29:00.600 So. These 553 00:29:00.600 --> 00:29:04.666 next few slides, I'm going to kind of walk you through their screen capture. 554 00:29:04.666 --> 00:29:06.100 So it's not going to look. 555 00:29:06.100 --> 00:29:08.200 It's not like I'm actually logged in and walking through. 556 00:29:08.200 --> 00:29:11.766 But when a technologist logs in to the software 557 00:29:12.366 --> 00:29:15.100 and tech side, is the company that developed the software, 558 00:29:15.100 --> 00:29:18.100 we clinically validated, but they actually developed the software. 559 00:29:18.466 --> 00:29:19.300 This is what they see. 560 00:29:19.300 --> 00:29:22.300 So the data like you see is always at the top. 561 00:29:22.666 --> 00:29:25.500 And then there's runs and you can go up here and click this 562 00:29:25.500 --> 00:29:27.000 and you can assign a run to you. 563 00:29:27.000 --> 00:29:29.066 So only you see the results. 564 00:29:29.066 --> 00:29:32.100 And then you just go through them one at a time looking at the images. 565 00:29:32.566 --> 00:29:35.366 And this is an example of of a positive. 566 00:29:35.366 --> 00:29:38.966 Now this was this was an older iteration of the model. 567 00:29:38.966 --> 00:29:41.166 So it doesn't look quite like this now. 568 00:29:41.166 --> 00:29:43.200 But you, you get the general idea. 569 00:29:43.200 --> 00:29:46.033 So this would be the exception. I blocked it out. 570 00:29:46.033 --> 00:29:48.466 And here's what the software found. 571 00:29:48.466 --> 00:29:50.600 It found 20 things that called blastocyst. 572 00:29:50.600 --> 00:29:54.233 This a couple things grew to 83 things day in to me. 573 00:29:54.233 --> 00:29:58.066 But 83 plus things and elements by the way. 574 00:29:58.066 --> 00:30:01.200 Now these are actually combined. 575 00:30:01.500 --> 00:30:04.066 So the technologist sees these as a call. 576 00:30:04.066 --> 00:30:05.533 It's not separated. 577 00:30:05.533 --> 00:30:10.700 We didn't want our technologist biasing the identification based on images. 578 00:30:12.300 --> 00:30:15.733 So, we actually combine those and the 579 00:30:15.733 --> 00:30:20.200 so the tech actually has to pull the slide to manually separate these. 580 00:30:20.200 --> 00:30:23.000 The reason there's a difference is a putative pathogen. 581 00:30:23.000 --> 00:30:27.066 So it's crucial that if it's defrag you accurately identified defrag 582 00:30:27.333 --> 00:30:28.800 and it flagged 11 troughs. 583 00:30:28.800 --> 00:30:30.966 So here's what it flagged as blast. 584 00:30:30.966 --> 00:30:33.366 So this this isn't too suspicious. 585 00:30:33.366 --> 00:30:34.666 This is kind of suspicious. 586 00:30:34.666 --> 00:30:36.566 Let's scroll down a little bit. 587 00:30:36.566 --> 00:30:40.933 Now we're getting into the defrag and the and alarm exploded me but 588 00:30:41.433 --> 00:30:43.600 I don't know to me, these are pretty suspicious 589 00:30:43.600 --> 00:30:47.900 for a protozoan trophies like this one, the carries, kind of looks fragmented. 590 00:30:47.900 --> 00:30:51.733 Maybe it's defrag, but they all only have one nucleus. 591 00:30:52.000 --> 00:30:55.333 So what the technologist can do is they can actually click on one, 592 00:30:55.966 --> 00:30:58.966 and it brings it up here where they can get a better view of it. 593 00:30:59.233 --> 00:31:01.866 Then they can click on this slide called exemplars. 594 00:31:01.866 --> 00:31:05.133 And that shows other images of indole, 595 00:31:05.133 --> 00:31:08.566 Imax and Iota amoeba that's in the training software. 596 00:31:08.566 --> 00:31:12.500 So they could see different stains, different fixative, so they could see 597 00:31:12.500 --> 00:31:16.566 all the different morphologic variation to say, hey, is this some 598 00:31:17.566 --> 00:31:19.733 is this legitimate or not? 599 00:31:19.733 --> 00:31:22.866 And if they think it is, they can then pull the slide for manual read. 600 00:31:22.866 --> 00:31:26.033 So in this case you have a technologist mark this 601 00:31:26.033 --> 00:31:29.033 and then the slide was pulled for manual read. 602 00:31:29.133 --> 00:31:30.733 A little background information. 603 00:31:30.733 --> 00:31:32.400 I was actually prepping these slot. 604 00:31:32.400 --> 00:31:35.166 When I prep these slides for an earlier talk. 605 00:31:35.166 --> 00:31:37.633 This slide had not been analyzed yet. 606 00:31:37.633 --> 00:31:40.133 Afterwards I went back and looked at the A session 607 00:31:40.133 --> 00:31:42.733 to see what it was reported ed as. 608 00:31:42.733 --> 00:31:46.266 And this indeed was I o to me, but I so 609 00:31:47.333 --> 00:31:50.333 I had you know, I had a pretty good idea that's what it was even before 610 00:31:50.333 --> 00:31:53.333 it was actually read by the technologist. 611 00:31:54.666 --> 00:31:57.033 Here's an example of a clean negative slide. 612 00:31:57.033 --> 00:31:59.366 And I made the images kind of small, 613 00:31:59.366 --> 00:32:02.866 just so you can get a full breadth of what the software flags. 614 00:32:03.266 --> 00:32:04.900 So when a technologist 615 00:32:04.900 --> 00:32:08.500 is going through and reading their slides, this is what a negative might look like. 616 00:32:08.833 --> 00:32:12.900 So you could still see that it flagged some stuff and calls it parasites. 617 00:32:12.900 --> 00:32:17.233 But a technologist looking at all these images can easily say 618 00:32:17.233 --> 00:32:18.533 that this is all garbage. 619 00:32:18.533 --> 00:32:24.000 And so you just went from spending two and a half to five minutes 620 00:32:24.433 --> 00:32:29.600 manually reading a trigram slide to spending 10s saying, okay, this one's 621 00:32:29.600 --> 00:32:33.366 negative, I can mark it is negative and move on to the next one. 622 00:32:35.300 --> 00:32:38.633 We also created prevalence regions. 623 00:32:38.633 --> 00:32:41.633 These are the prevalence regions I enlarged this one. 624 00:32:41.700 --> 00:32:43.766 So you can get a good view of it. 625 00:32:43.766 --> 00:32:45.100 The prevalence regions. 626 00:32:45.100 --> 00:32:48.100 And there should be 11 of them on a complete scan. 627 00:32:48.166 --> 00:32:51.900 The prevalence regions were designed to simulate what it would look like 628 00:32:51.900 --> 00:32:55.800 if you were actually looking through the microscope at 100 X with oil. 629 00:32:56.100 --> 00:32:59.533 And we designed this so the technologist can use these images 630 00:32:59.533 --> 00:33:03.600 for a enumerating red blood cells and white blood cells. 631 00:33:03.866 --> 00:33:06.700 So they didn't actually have to pull the slide to do that. 632 00:33:06.700 --> 00:33:10.600 We thought this would be just if we could just eliminate one more aspect 633 00:33:10.600 --> 00:33:14.566 of the technologist having to review, slides. 634 00:33:15.466 --> 00:33:16.566 That's what they would do. 635 00:33:16.566 --> 00:33:19.766 So if this was a high level red cell, white cell 636 00:33:19.766 --> 00:33:22.766 slide, they could say looking at this field, oh man. 637 00:33:22.800 --> 00:33:25.200 There's like it's not on the slide. 638 00:33:25.200 --> 00:33:26.533 I'm just using this as an example. 639 00:33:26.533 --> 00:33:29.100 Oh man. There's like 150 white cells here. 640 00:33:29.100 --> 00:33:30.900 This is probably going to be four plus. 641 00:33:30.900 --> 00:33:33.733 So that's the way that's why we designed this. 642 00:33:33.733 --> 00:33:36.900 In theory you could also use it for enumerating blastocysts. 643 00:33:36.900 --> 00:33:41.600 But I prefer to do that from the manual smear just because some docs 644 00:33:41.600 --> 00:33:45.600 only treat blastocyst this if it's at the three and four plus level. 645 00:33:45.600 --> 00:33:47.233 So I want to make sure 646 00:33:47.233 --> 00:33:51.133 that the blast still counts are or none of them are accurate counts. 647 00:33:51.133 --> 00:33:54.766 But I want to make sure that the blast blastoma enumerations 648 00:33:54.766 --> 00:33:58.700 are relatively realistic to what's actually seen on the slide. 649 00:34:01.066 --> 00:34:01.300 These 650 00:34:01.300 --> 00:34:04.333 next few models are from the modified assets arsenal. 651 00:34:04.333 --> 00:34:07.533 These are models from training, not the actual lab. 652 00:34:07.533 --> 00:34:11.733 So they look a little different than what the technologist actually sees. 653 00:34:12.100 --> 00:34:14.900 But I just wanted you to see what cryptosporidium 654 00:34:14.900 --> 00:34:17.933 stain with modified acid fest looks like. 655 00:34:17.933 --> 00:34:20.066 Look, you can even see the individual spores on. 656 00:34:20.066 --> 00:34:23.066 It's on some of these. 657 00:34:23.233 --> 00:34:27.233 And if you click on it again you can see you can see a closeup. 658 00:34:27.233 --> 00:34:29.266 You can really see those in spores. 659 00:34:29.266 --> 00:34:30.700 Oh right inside the assist. 660 00:34:30.700 --> 00:34:34.533 And again, if the technologist wasn't sure, they could click on examples 661 00:34:34.533 --> 00:34:37.533 to get an idea of what it should look like. 662 00:34:37.733 --> 00:34:40.033 Here's, cryptosporidium. 663 00:34:40.033 --> 00:34:41.700 Excuse me. Here, Cyclospora. 664 00:34:41.700 --> 00:34:44.700 As many of you who stay in Cyclospora know, 665 00:34:45.500 --> 00:34:47.300 they don't often like to stain. 666 00:34:47.300 --> 00:34:48.733 So here are some stain versions. 667 00:34:48.733 --> 00:34:51.600 And then you have these ghost versions. 668 00:34:51.600 --> 00:34:55.800 And, by the way, you guys know what you call a whole bunch of these 669 00:34:55.800 --> 00:35:00.300 round things that look like ohs a bunch of ghost forms in a row. Ooh, 670 00:35:01.400 --> 00:35:01.866 sorry. 671 00:35:01.866 --> 00:35:04.866 Bad joke. 672 00:35:05.666 --> 00:35:08.366 So here's our lab workflow right now. 673 00:35:08.366 --> 00:35:11.233 So you start with, 674 00:35:11.233 --> 00:35:14.166 I, I've already kind of gone over this, but here's in kind of a flowchart method. 675 00:35:14.166 --> 00:35:17.133 So this is just for the lab image analysis. 676 00:35:17.133 --> 00:35:20.433 You have a suspect positive organism, white cells. 677 00:35:21.133 --> 00:35:24.000 No. You can report the tri chroma is negative 678 00:35:24.000 --> 00:35:27.000 without manually reading the slide. Yes. 679 00:35:27.200 --> 00:35:30.600 Flag exemplars add comments appropriate and calculate 680 00:35:30.600 --> 00:35:32.800 the red cells and white cells. 681 00:35:32.800 --> 00:35:36.866 Based on the prevalence region, the technologist reviews the slide 682 00:35:37.166 --> 00:35:42.033 where the organisms flag by the software found on manually read. 683 00:35:42.400 --> 00:35:45.400 If yes, you can report as possible or excuse me 684 00:35:45.400 --> 00:35:48.600 report is positive and calculate the blastocyst. 685 00:35:48.600 --> 00:35:50.800 This is needed per your SOP. 686 00:35:50.800 --> 00:35:55.033 If you if the organisms weren't found on the smear, you could direct it 687 00:35:55.033 --> 00:36:00.066 to either a medical director or specialist review to analyze the images in our lab. 688 00:36:00.066 --> 00:36:03.000 We let the medical director do this because of the possible treatment 689 00:36:03.000 --> 00:36:07.366 ramifications of calling a positive parasite based on just images. 690 00:36:07.633 --> 00:36:10.533 So right now that's handled just by the medical director. 691 00:36:10.533 --> 00:36:11.666 They review the images. 692 00:36:11.666 --> 00:36:13.433 Do they say it's positive? Yep. 693 00:36:13.433 --> 00:36:14.800 Report is positive. 694 00:36:14.800 --> 00:36:16.466 Nope. Report is negative. 695 00:36:18.066 --> 00:36:21.066 And with that here's a really fun example. 696 00:36:21.066 --> 00:36:22.600 This here was the clinical. 697 00:36:22.600 --> 00:36:24.266 This is what the software found. 698 00:36:24.266 --> 00:36:25.433 It called a couple blasts. 699 00:36:25.433 --> 00:36:28.100 So it called this thing giardia. 700 00:36:28.100 --> 00:36:32.000 I call this junk into me but I enlarged the giardia here. 701 00:36:32.533 --> 00:36:35.233 In my opinion, this is pretty textbook giardia. 702 00:36:35.233 --> 00:36:38.333 You can see the nuclei, you can kind of see the sucking disk. 703 00:36:38.700 --> 00:36:42.366 This might be a several or even a flagella coming off the bottom. 704 00:36:42.566 --> 00:36:44.733 This might be a flagella up here, 705 00:36:44.733 --> 00:36:47.400 but the technologist spent like 20 minutes reading this slide 706 00:36:47.400 --> 00:36:50.400 and could not find this one stinking giardia trough. 707 00:36:50.433 --> 00:36:52.600 So the question is. 708 00:36:52.600 --> 00:36:54.833 So the medical director looked at the images and said, yep, 709 00:36:54.833 --> 00:36:57.833 he gave us thumbs up and said, okay, you could report this is great. 710 00:36:58.200 --> 00:37:02.266 But the question is, is without the software, would we have missed this? 711 00:37:02.800 --> 00:37:06.600 Our positivity rate has gone up from about two 712 00:37:06.600 --> 00:37:10.133 and a half to 4% since we've initiated the software. 713 00:37:10.366 --> 00:37:12.300 Now there's a lot of other things that can affect that 714 00:37:12.300 --> 00:37:14.400 because we're a large reference lab. 715 00:37:14.400 --> 00:37:18.033 You know, we may have got new clients that have a higher positivity rate. 716 00:37:18.333 --> 00:37:23.633 I don't know, but you got, the general feeling of the technologists in the lab 717 00:37:23.966 --> 00:37:25.966 is that this software is finding a lot of things 718 00:37:25.966 --> 00:37:27.666 that they might have otherwise missed. 719 00:37:28.733 --> 00:37:29.633 So I threw in 720 00:37:29.633 --> 00:37:32.666 some fun slides for future development. 721 00:37:32.666 --> 00:37:36.500 So we're currently working on the wet mount for this. 722 00:37:36.500 --> 00:37:40.066 And when this is finally clinically validated, our entire open 723 00:37:40.066 --> 00:37:43.266 parasite workflow is going to be prescreened by I. 724 00:37:44.133 --> 00:37:47.800 Now the web mount is a heck of a lot more challenging. 725 00:37:47.800 --> 00:37:52.700 So we had a lot of challenges that we had to overcome for the wet mount model. 726 00:37:53.800 --> 00:37:56.733 First of all, there's more classes to consider. 727 00:37:56.733 --> 00:37:59.466 So with the wet mount, you now have the concern, 728 00:37:59.466 --> 00:38:02.233 concern yourself with all those pesky worms 729 00:38:02.233 --> 00:38:05.233 that people might see like three times in their career. 730 00:38:05.266 --> 00:38:08.933 So you had to test on just this on the Mount Sinai. 731 00:38:09.266 --> 00:38:11.833 You had to test on, 732 00:38:11.833 --> 00:38:16.033 you know, high rodent Olympus Nano, which is the new name for hymen leopards. 733 00:38:16.033 --> 00:38:17.966 Nana. You know, you had to test. 734 00:38:17.966 --> 00:38:20.466 You had to test on all these kind of esoteric worms. 735 00:38:20.466 --> 00:38:23.866 So, even those that are not endemic to North America, 736 00:38:23.866 --> 00:38:26.600 because maybe that specimen that came through your lab, 737 00:38:26.600 --> 00:38:30.900 maybe it's in an immigrant or a refugee, they could have some tropical worm 738 00:38:30.900 --> 00:38:33.100 and you have to be able to account for that. 739 00:38:33.100 --> 00:38:34.733 So how do we get that? 740 00:38:34.733 --> 00:38:38.100 Well, we used leftover proficiency specimens. 741 00:38:38.433 --> 00:38:40.633 We use like luckily we're a large reference lab. 742 00:38:40.633 --> 00:38:42.700 So we get a lot of the common helminths. 743 00:38:42.700 --> 00:38:45.066 We also had to reach out to international collaborators. 744 00:38:45.066 --> 00:38:48.566 So we worked with people in Africa, in the Philippines to get us things 745 00:38:48.566 --> 00:38:53.700 like just awesome amounts and I just this on the japonica, capillary Philippines's 746 00:38:55.566 --> 00:38:58.566 you have to develop a mounting media to slow the drying. 747 00:38:58.733 --> 00:39:01.566 If you're going to scan 20 wet mounts at once, 748 00:39:01.566 --> 00:39:03.333 you got to make sure that while it's scanning 749 00:39:03.333 --> 00:39:06.766 the first couple slides, your last slides in the run aren't drying. 750 00:39:07.100 --> 00:39:11.500 So I had to develop, a mounting medium that, 751 00:39:12.433 --> 00:39:16.200 and we had to do some drying studies to see how long they were last. We. 752 00:39:16.400 --> 00:39:18.233 I can't tell you exactly what that is yet, 753 00:39:18.233 --> 00:39:20.666 because it's still technically proprietary. 754 00:39:20.666 --> 00:39:23.166 Until we're done and we can publish the data, 755 00:39:23.166 --> 00:39:25.533 but it's nothing really special, is spectacular. 756 00:39:25.533 --> 00:39:28.533 And probably anyone would come up with the same thing on their own. 757 00:39:28.600 --> 00:39:31.233 But I can't really go into the the details of it just yet. 758 00:39:31.233 --> 00:39:33.766 I'm sorry. 759 00:39:33.766 --> 00:39:37.066 You need to have a scanner that loads the slides flat. 760 00:39:37.366 --> 00:39:40.833 The scanner we clinically validated for tri chrome and modified acid fast. 761 00:39:41.166 --> 00:39:43.233 Interestingly, the slides come in at an angle, 762 00:39:43.233 --> 00:39:45.333 so you can't use that as a wet mount. 763 00:39:45.333 --> 00:39:48.633 So you need something where the the slides not only come in flat, 764 00:39:49.166 --> 00:39:52.400 but you have to make sure the cover slide doesn't shift. 765 00:39:52.733 --> 00:39:55.533 And this is also plays in the role of when you're making your smears. 766 00:39:55.533 --> 00:39:59.033 You want to develop a template that you can make your slides on. 767 00:39:59.300 --> 00:40:03.900 So your cover slip is always in the same place or as best you can. 768 00:40:04.433 --> 00:40:06.933 And then of course development of the scan area 769 00:40:06.933 --> 00:40:10.266 and the tri chrome a technologist is expected to read 770 00:40:10.800 --> 00:40:15.166 100 fields at 1100 x with oil, but in a wet mound. 771 00:40:15.166 --> 00:40:17.700 Remember, in your typical wet mount workflow, 772 00:40:17.700 --> 00:40:20.833 you have to read the entire 22 by 22 cover slip. 773 00:40:21.100 --> 00:40:24.900 So you want to get a scan area that covers as much as that as possible. 774 00:40:25.200 --> 00:40:28.866 In case you've got one of those worms that sheds eggs and really low numbers. 775 00:40:28.866 --> 00:40:32.666 And there's only one word make on the slide, you want to make sure you find it, 776 00:40:33.866 --> 00:40:35.866 and finding optimal levels to scan. 777 00:40:35.866 --> 00:40:39.666 One thing we learned is that the protozoan cysts in the egg 778 00:40:39.666 --> 00:40:42.666 and the helminth eggs actually settle in different places. 779 00:40:42.933 --> 00:40:47.533 So we actually have to scan to we have to scan the whole cover slip twice. 780 00:40:47.833 --> 00:40:50.000 So that goes into your scan time. 781 00:40:50.000 --> 00:40:52.066 You got to be able to scan two levels 782 00:40:52.066 --> 00:40:55.400 at an efficient time, which ideally is five minutes or less. 783 00:40:55.900 --> 00:40:58.900 Luckily, we've overcome nearly all of these. 784 00:41:02.500 --> 00:41:04.733 So what classes did we want to train? 785 00:41:04.733 --> 00:41:07.733 As I said, remember, you got to cover 786 00:41:07.800 --> 00:41:11.233 almost all potential parasites that a patient may have. 787 00:41:11.500 --> 00:41:15.633 So for the protozoa, there's a lot of crossover with, 788 00:41:16.933 --> 00:41:20.900 the tri chrome, except in many cases you can see it's also there cysts. 789 00:41:21.300 --> 00:41:24.633 We did train and Tommy Bassett troughs and troughs 790 00:41:24.633 --> 00:41:26.700 because they're fairly distinctive. 791 00:41:26.700 --> 00:41:32.000 But we also had to create a a class that captures all those random 792 00:41:32.000 --> 00:41:36.800 troves of small, protozoa that on a wet mount are all going to look the same. 793 00:41:37.000 --> 00:41:39.133 I don't all of the ideally, of course, you know, 794 00:41:39.133 --> 00:41:41.700 you don't want to use a wet mount for detecting these things, 795 00:41:41.700 --> 00:41:43.200 but the software is going to find them. 796 00:41:43.200 --> 00:41:46.500 So we just kind of created this dumping ground for all these 797 00:41:46.666 --> 00:41:49.566 ideally that the tri Chrome software 798 00:41:49.566 --> 00:41:52.933 will flag all these and catch them because you have that balance. 799 00:41:53.366 --> 00:41:54.333 But we also added, 800 00:41:55.500 --> 00:41:58.733 is coli that's the proper name for stadium coli. 801 00:41:59.100 --> 00:42:03.000 We added Cyclospora, even though typically you don't want to use Cyclospora 802 00:42:03.000 --> 00:42:04.133 for a web. 803 00:42:04.133 --> 00:42:06.600 Now for Cyclospora, it trains very easily. 804 00:42:06.600 --> 00:42:09.600 And we added Cisco sauce probably to this one. 805 00:42:10.366 --> 00:42:13.866 And then the worm classes, what worm classes did we want to train? 806 00:42:13.866 --> 00:42:16.200 Well, obviously you got to hit the big ones. 807 00:42:16.200 --> 00:42:18.766 Ask was trickier is hookworm strategy. 808 00:42:18.766 --> 00:42:21.433 Pin worms are not commonly detected at wet mounts. 809 00:42:21.433 --> 00:42:24.100 But we're makers are very easy to train. 810 00:42:24.100 --> 00:42:28.200 We did we were able to acquire para capillary, Philippine fences. 811 00:42:28.200 --> 00:42:31.866 That's the proper name for capillary Philippines's tinea eggs. 812 00:42:32.233 --> 00:42:35.533 The hymen lipid says those fish tapeworm eggs. 813 00:42:35.533 --> 00:42:38.200 This is your formal die Philip bathroom complex. 814 00:42:38.200 --> 00:42:39.933 There's now three genera. 815 00:42:39.933 --> 00:42:42.933 And then we felt that these worms really? 816 00:42:43.500 --> 00:42:45.633 There were some other worms we wanted we couldn't get, 817 00:42:45.633 --> 00:42:47.900 but we felt that these ones were essential. 818 00:42:47.900 --> 00:42:49.333 So we trained on just the soma. 819 00:42:49.333 --> 00:42:51.133 Manzanar is just the soma japonica. 820 00:42:52.533 --> 00:42:52.900 Cleaner. 821 00:42:52.900 --> 00:42:55.833 Orcas and orcas, Paragon, amiss. 822 00:42:55.833 --> 00:42:59.333 And I have an asterisk here because this is the last class 823 00:42:59.333 --> 00:43:01.166 we're having trouble finding. 824 00:43:01.166 --> 00:43:04.400 Once we get it trained on these, either or both of these, 825 00:43:04.800 --> 00:43:07.800 then we can go into clinical validation. 826 00:43:07.933 --> 00:43:11.300 So for my last few slides, I'll show you some of our, 827 00:43:12.433 --> 00:43:14.033 some of the things we see in a wet mouth. 828 00:43:14.033 --> 00:43:16.733 These first few images weren't taken with the scanner. 829 00:43:16.733 --> 00:43:20.866 These were my drawing experiments for my, the mounting media. 830 00:43:20.866 --> 00:43:23.033 So this was actually taken after like two hours 831 00:43:23.033 --> 00:43:26.100 of sitting at room temp with this modified mounting media. 832 00:43:26.100 --> 00:43:30.266 Here's some killer mouse sticks, Giardia, a trough in a cyst. 833 00:43:31.400 --> 00:43:33.600 Here's a strongyloides larva, 834 00:43:33.600 --> 00:43:36.266 here's a US egg, 835 00:43:36.266 --> 00:43:39.066 here's a balloon to your 80s trough. 836 00:43:39.066 --> 00:43:43.200 So that was for the, that was for the, the drying challenges 837 00:43:43.200 --> 00:43:44.766 for the mounting media 838 00:43:44.766 --> 00:43:47.700 and now I'm going to show you and you guys are getting a sneak peak 839 00:43:47.700 --> 00:43:51.966 because this isn't validated yet of what the model finds. 840 00:43:52.600 --> 00:43:54.333 So here's paragon of mistakes, 841 00:43:55.833 --> 00:43:57.900 here's strongyloides larvae. 842 00:43:57.900 --> 00:44:01.000 So because strongyloides can be coiled up, what we did is 843 00:44:01.000 --> 00:44:04.000 we only trained the anterior third of the worm. 844 00:44:04.066 --> 00:44:07.500 So we trained from the mouthparts to the bulbous esophagus. 845 00:44:07.766 --> 00:44:10.766 So this is what the software focuses on. 846 00:44:11.733 --> 00:44:13.800 Here's some hookworm eggs. 847 00:44:13.800 --> 00:44:16.800 Here's some orcas eggs. 848 00:44:18.700 --> 00:44:21.700 Here's pin worm eggs. 849 00:44:21.766 --> 00:44:23.466 And here's a bunch of our assays. 850 00:44:23.466 --> 00:44:24.400 We trained. 851 00:44:24.400 --> 00:44:28.366 We trained fertile, immolated, fertile decor dictated, infertile, 852 00:44:28.366 --> 00:44:31.900 immolated, infertile decor mated as four separate classes. 853 00:44:32.400 --> 00:44:34.866 But the end user is only going to see our scorers. 854 00:44:34.866 --> 00:44:37.900 I've never seen a clinical ask or a specimen that didn't have fertile 855 00:44:37.900 --> 00:44:40.900 mammalian eggs, so that's going to be the most important. 856 00:44:41.700 --> 00:44:44.700 Here's Berlanti or these collide troughs. 857 00:44:45.066 --> 00:44:46.600 Here's Cisco I saw spur belly. 858 00:44:46.600 --> 00:44:47.766 Oh cysts. 859 00:44:47.766 --> 00:44:49.800 You know, this might be overlooked 860 00:44:49.800 --> 00:44:52.800 in a normal manual read, but the software can find it. 861 00:44:53.233 --> 00:44:55.633 Here's some shows this on my Mount Sinai eggs. 862 00:44:55.633 --> 00:44:57.700 This one has a lateral spine. 863 00:44:57.700 --> 00:45:00.766 This one the spine is probably above or below the egg, 864 00:45:01.100 --> 00:45:05.233 but the software can recognize egg morphology even without the spine. 865 00:45:06.333 --> 00:45:07.533 So in 866 00:45:07.533 --> 00:45:10.800 closing in overall, for the whole sake of this whole talk, 867 00:45:11.400 --> 00:45:15.933 artificial intelligence, such as machine learning, can improve the workflow 868 00:45:16.266 --> 00:45:21.000 in a diagnostic parasitology lab by improving turnaround time, 869 00:45:21.000 --> 00:45:24.266 by reducing the amount of time needed to manually read slides, 870 00:45:24.866 --> 00:45:28.800 ideally screened out at least 70 to 80% of your negative specimens, 871 00:45:29.233 --> 00:45:33.566 possibly reduce ergonomic injuries associated with prolonged microscopy. 872 00:45:34.200 --> 00:45:36.000 Improve employee satisfaction. 873 00:45:36.000 --> 00:45:39.400 So instead of looking at negative poops all day, your employees, 874 00:45:39.400 --> 00:45:42.633 whether the slides are actually manually back reading, are positive. 875 00:45:42.900 --> 00:45:45.800 They're spending more time looking at a positive slide. 876 00:45:45.800 --> 00:45:47.033 And then therefore they're actually 877 00:45:47.033 --> 00:45:50.433 increasing their competency for identifying organisms. 878 00:45:51.066 --> 00:45:54.466 It might be more engaging for an increasingly younger workforce. 879 00:45:54.466 --> 00:45:59.866 And this last one there's a little caveat allows for image analysis remotely. 880 00:45:59.866 --> 00:46:00.100 You know, 881 00:46:00.100 --> 00:46:03.533 since the Covid pandemic, a lot more people are finding ways to work from home, 882 00:46:03.966 --> 00:46:08.400 maybe someone collagen from home and do all the software analysis from home. 883 00:46:08.733 --> 00:46:11.733 But after I added this to the slide, I started thinking, 884 00:46:12.300 --> 00:46:15.166 technically, you're only allowed to analyze a specimen 885 00:46:15.166 --> 00:46:17.833 in a building that's been clear, certified. 886 00:46:17.833 --> 00:46:21.133 So I don't know if you can actually do remote analysis. 887 00:46:21.233 --> 00:46:23.766 You could do an analysis remotely. 888 00:46:23.766 --> 00:46:28.300 You might be able to flag the suspect positives, but you definitely couldn't 889 00:46:28.300 --> 00:46:32.100 do formal reporting unless you had your your home clear lysed. 890 00:46:32.400 --> 00:46:35.800 And I don't think you guys want clear inspectors come in CMS inspectors 891 00:46:35.800 --> 00:46:39.166 come into your lab or your home every every two years. 892 00:46:39.166 --> 00:46:40.433 Right. So 893 00:46:41.533 --> 00:46:44.533 anyway, that wraps up my talk. 894 00:46:45.300 --> 00:46:47.500 And there's a lot of people I'm not going to rattle off 895 00:46:47.500 --> 00:46:50.566 names, but an endeavor like this couldn't be done 896 00:46:50.566 --> 00:46:55.933 without a coordinated effort of our medical directorship, R&D. 897 00:46:56.400 --> 00:46:59.966 Some of these people were very involved in the scanning processing. 898 00:47:00.100 --> 00:47:04.200 Some are involved into some of the, the early development ideas 899 00:47:04.566 --> 00:47:09.100 and of course, all our folks at tech side, the engineers who helped develop 900 00:47:09.100 --> 00:47:12.933 the technology, who have developed the software, 901 00:47:13.200 --> 00:47:18.033 analyze the software and help come up with the best models for clinical practice. 902 00:47:18.033 --> 00:47:21.900 So it was a really huge endeavor, and it couldn't really be done without 903 00:47:22.233 --> 00:47:25.933 a lot of great contributions from numerous different people at all levels.