1 00:00:08,000 --> 00:00:10,719 Hello, and welcome to this episode of the 2 00:00:10,719 --> 00:00:12,660 Physics World weekly podcast, 3 00:00:12,960 --> 00:00:14,099 which is sponsored 4 00:00:14,400 --> 00:00:15,779 by Sun nuclear. 5 00:00:16,774 --> 00:00:18,795 Our guest is the medical physicist, 6 00:00:19,175 --> 00:00:20,155 Todd McNutt, 7 00:00:20,535 --> 00:00:21,675 who's in conversation 8 00:00:22,054 --> 00:00:24,155 with Physics World's Tammy Freeman. 9 00:00:24,935 --> 00:00:26,875 They explore how an artificial 10 00:00:27,255 --> 00:00:29,035 intelligence based tool 11 00:00:29,414 --> 00:00:32,549 can help improve the quality of radiation 12 00:00:32,929 --> 00:00:35,510 therapy plans for cancer treatments. 13 00:00:36,530 --> 00:00:39,190 But first, a message from our sponsor. 14 00:00:40,289 --> 00:00:43,030 Treatment planning in radiation oncology 15 00:00:43,890 --> 00:00:44,390 requires 16 00:00:44,770 --> 00:00:46,390 balancing plan quality, 17 00:00:47,065 --> 00:00:47,565 efficiency, 18 00:00:48,024 --> 00:00:49,325 and clinical judgment. 19 00:00:50,104 --> 00:00:51,245 Plan AI, 20 00:00:51,704 --> 00:00:53,804 a sun nuclear product, 21 00:00:54,344 --> 00:00:57,565 supports that process by using large scale, 22 00:00:58,104 --> 00:00:59,725 real world radiotherapy 23 00:01:00,185 --> 00:01:00,685 data 24 00:01:01,179 --> 00:01:02,479 to predict achievable 25 00:01:02,859 --> 00:01:03,359 DVHs 26 00:01:04,219 --> 00:01:06,319 for organ at risk sparing 27 00:01:06,939 --> 00:01:08,319 and tumor coverage 28 00:01:08,700 --> 00:01:09,840 for each patient. 29 00:01:10,620 --> 00:01:11,359 By providing 30 00:01:11,739 --> 00:01:14,560 data driven planning objectives upfront, 31 00:01:15,295 --> 00:01:19,075 plan AI helps reduce trial and error optimization, 32 00:01:20,174 --> 00:01:21,314 improve consistency 33 00:01:21,694 --> 00:01:22,594 across planners, 34 00:01:22,974 --> 00:01:24,354 and support confident 35 00:01:24,734 --> 00:01:26,274 clinical decision making. 36 00:01:27,390 --> 00:01:31,409 Integrated directly into the treatment planning workflow, 37 00:01:31,950 --> 00:01:32,450 Plan 38 00:01:32,990 --> 00:01:35,329 AI is designed to help teams deliver 39 00:01:35,790 --> 00:01:37,170 high quality plans 40 00:01:37,710 --> 00:01:38,210 efficiently, 41 00:01:38,829 --> 00:01:39,329 consistently, 42 00:01:39,870 --> 00:01:42,609 and with patient care at the center. 43 00:01:50,814 --> 00:01:53,935 Artificial intelligence is a valuable tool used within 44 00:01:53,935 --> 00:01:57,310 many areas of health care, from analyzing diagnostic 45 00:01:57,450 --> 00:01:58,349 medical images 46 00:01:58,729 --> 00:02:01,709 to empowering the individuals that plan cancer treatments 47 00:02:01,769 --> 00:02:03,229 using radiation therapy. 48 00:02:03,930 --> 00:02:06,829 Today, we're gonna be talking about Plan AI, 49 00:02:07,290 --> 00:02:10,425 an AI powered plan quality improvement tool. 50 00:02:11,224 --> 00:02:14,585 Plan AI uses an extensive database of thousands 51 00:02:14,585 --> 00:02:16,685 of previous radiotherapy treatments 52 00:02:17,064 --> 00:02:19,705 to predict the lowest possible radiation dose to 53 00:02:19,705 --> 00:02:21,965 healthy tissues for each new patient. 54 00:02:22,824 --> 00:02:24,925 Treatment planners then use this information 55 00:02:25,370 --> 00:02:28,169 to define goals that streamline and automate the 56 00:02:28,169 --> 00:02:30,430 creation of a best achievable plan. 57 00:02:31,530 --> 00:02:34,090 In addition to discussing the benefits that Plan 58 00:02:34,090 --> 00:02:35,709 AI brings for radiotherapy 59 00:02:36,009 --> 00:02:38,030 patients and cancer treatment centers, 60 00:02:38,525 --> 00:02:40,284 We'll also take a look at the platform's 61 00:02:40,284 --> 00:02:40,784 evolution 62 00:02:41,724 --> 00:02:43,745 from an idea developed by OncoSpace, 63 00:02:44,365 --> 00:02:47,425 an academic collaboration at Johns Hopkins University, 64 00:02:48,284 --> 00:02:50,925 to a clinical product offered today by Sun 65 00:02:50,925 --> 00:02:52,740 nuclear, a US manufacturer 66 00:02:53,040 --> 00:02:54,980 of radiation equipment and software. 67 00:02:56,159 --> 00:02:58,500 I'm pleased to be joined by Todd McNutt, 68 00:02:58,879 --> 00:03:02,319 associate professor of radiation oncology physics at Johns 69 00:03:02,319 --> 00:03:03,379 Hopkins University 70 00:03:03,935 --> 00:03:05,394 and the founder of OncoSpace. 71 00:03:06,254 --> 00:03:08,034 Welcome to the podcast, Todd. 72 00:03:08,574 --> 00:03:10,514 Why, thank you for having me. 73 00:03:11,055 --> 00:03:13,294 So first of all, can you share the 74 00:03:13,294 --> 00:03:14,995 story of how the OncoSpace 75 00:03:15,294 --> 00:03:15,794 project 76 00:03:16,110 --> 00:03:18,209 began back in 2007? 77 00:03:19,549 --> 00:03:20,349 Sure. Yeah. We, 78 00:03:21,389 --> 00:03:22,669 back in 2007 79 00:03:22,669 --> 00:03:23,870 is when we really had, 80 00:03:24,830 --> 00:03:27,810 initiatives to go electronic in the clinical environment. 81 00:03:28,430 --> 00:03:29,889 And there were several groups 82 00:03:31,074 --> 00:03:32,935 discussing how we could better use, 83 00:03:33,395 --> 00:03:36,294 clinical data for discovery and knowledge generation. 84 00:03:37,074 --> 00:03:38,435 And I was in charge with, 85 00:03:39,155 --> 00:03:41,574 moving our clinic to an electronic platform, 86 00:03:41,969 --> 00:03:44,530 And we had several meetings with with folks 87 00:03:44,530 --> 00:03:47,009 at Johns Hopkins, and one in particular was 88 00:03:47,009 --> 00:03:47,830 Alex Saleh, 89 00:03:48,449 --> 00:03:51,990 who did, the Sloan Digital Sky Surveys SkyServer 90 00:03:52,129 --> 00:03:54,870 project where he had built a large database 91 00:03:55,090 --> 00:03:58,155 of galaxies and stars and cosmos. And it 92 00:03:58,155 --> 00:03:59,294 became a huge, 93 00:03:59,835 --> 00:04:01,375 research platform where 94 00:04:01,675 --> 00:04:03,375 people could be junior astronomers 95 00:04:03,754 --> 00:04:05,294 using this data. And 96 00:04:05,675 --> 00:04:07,534 from that, there were other initiatives, 97 00:04:08,074 --> 00:04:10,314 like the in health with Scott Seeger, and 98 00:04:10,314 --> 00:04:11,675 and we had a lot of high level 99 00:04:11,675 --> 00:04:12,569 meetings. And 100 00:04:13,050 --> 00:04:13,550 I 101 00:04:13,849 --> 00:04:14,250 I 102 00:04:14,650 --> 00:04:16,889 taking from that, we really looked at going 103 00:04:16,889 --> 00:04:18,189 electronic and moving 104 00:04:18,490 --> 00:04:21,610 towards structured data collection for patients in the 105 00:04:21,610 --> 00:04:22,589 clinical environment 106 00:04:23,210 --> 00:04:26,490 and marrying that data with the treatment plans, 107 00:04:26,490 --> 00:04:28,029 our radiation dose delivery 108 00:04:28,524 --> 00:04:30,925 treatment plans in a database so that we 109 00:04:30,925 --> 00:04:33,904 could really start looking at how our, 110 00:04:34,685 --> 00:04:37,884 radiation dose distributions across the anatomy would affect 111 00:04:37,884 --> 00:04:38,865 patient outcomes, 112 00:04:39,644 --> 00:04:41,564 and things like that. And we just really 113 00:04:41,564 --> 00:04:43,680 looked at it like that and and took 114 00:04:43,680 --> 00:04:45,459 that opportunity to build it. 115 00:04:46,800 --> 00:04:49,600 Well, that's really interesting that it started from 116 00:04:49,600 --> 00:04:52,240 from the astronomy database and then moving on 117 00:04:52,240 --> 00:04:52,980 to to 118 00:04:53,360 --> 00:04:55,139 the idea of using this for radiotherapy. 119 00:04:56,000 --> 00:04:56,500 So 120 00:04:57,285 --> 00:05:00,404 moving further ahead, what inspired the transition from 121 00:05:00,404 --> 00:05:01,865 this academic research 122 00:05:02,404 --> 00:05:05,064 to founding the company OncoSpace Inc 123 00:05:05,365 --> 00:05:06,504 in 2019, 124 00:05:07,285 --> 00:05:09,764 which was then acquired by Sun Nuclear earlier 125 00:05:09,764 --> 00:05:10,504 this year? 126 00:05:11,125 --> 00:05:13,279 Mhmm. Yeah. So one of, one of the 127 00:05:13,279 --> 00:05:15,199 earlier things we were able to do with 128 00:05:15,199 --> 00:05:17,539 the with the database after we had populated 129 00:05:17,599 --> 00:05:18,180 it with, 130 00:05:19,120 --> 00:05:19,939 many patients 131 00:05:20,479 --> 00:05:21,379 is we were 132 00:05:21,680 --> 00:05:24,159 we had a project with, Russ Taylor and 133 00:05:24,159 --> 00:05:24,899 some students 134 00:05:25,454 --> 00:05:26,915 where we looked at, 135 00:05:27,774 --> 00:05:28,915 how what 136 00:05:29,294 --> 00:05:32,995 what features of patient's anatomy and target volumes 137 00:05:33,774 --> 00:05:35,634 impact our ability to generate 138 00:05:36,574 --> 00:05:38,975 a treatment plan to deliver dose that best 139 00:05:38,975 --> 00:05:39,475 avoids 140 00:05:40,060 --> 00:05:42,620 normal tissues while treating as best as possible 141 00:05:42,620 --> 00:05:44,779 the target binds. And we came up with 142 00:05:44,779 --> 00:05:47,120 a feature set that looked at the relationships 143 00:05:47,500 --> 00:05:48,000 between 144 00:05:49,019 --> 00:05:52,000 normal anatomy and targets and target complexity. 145 00:05:52,779 --> 00:05:55,965 And our earlier work allowed us to sort 146 00:05:55,965 --> 00:05:56,785 of predict 147 00:05:57,485 --> 00:05:59,905 that the expected doses from, 148 00:06:00,605 --> 00:06:03,485 from these shape relationship features, and it worked 149 00:06:03,485 --> 00:06:05,004 quite well. And we published on that a 150 00:06:05,004 --> 00:06:07,165 few times. And so at that point, we 151 00:06:07,165 --> 00:06:07,985 really knew 152 00:06:08,790 --> 00:06:11,430 we could tap into this data and and 153 00:06:11,430 --> 00:06:14,069 offer up this prediction that could help do 154 00:06:14,069 --> 00:06:15,750 treatment plans on a new patient. It was 155 00:06:15,750 --> 00:06:17,110 sort of at that point, we thought of 156 00:06:17,110 --> 00:06:18,810 it as personalized medicine. So 157 00:06:19,269 --> 00:06:21,589 what for this specific patient, how good of 158 00:06:21,589 --> 00:06:23,110 a treatment plan should I be able to 159 00:06:23,110 --> 00:06:25,024 create? And for the first time, we were 160 00:06:25,024 --> 00:06:26,084 able to see that. 161 00:06:26,464 --> 00:06:28,064 And I I always thought that that was 162 00:06:28,064 --> 00:06:30,865 something that was useful commercially, and we could 163 00:06:30,865 --> 00:06:33,264 get it out to other clinics and allow 164 00:06:33,264 --> 00:06:35,185 them to use it. But I didn't have, 165 00:06:35,185 --> 00:06:37,519 like, the strength or the wherewithal to do 166 00:06:37,519 --> 00:06:40,639 that back then. And what it I think 167 00:06:40,639 --> 00:06:42,399 what it really took is my my friend 168 00:06:42,399 --> 00:06:45,360 Praveen Sinha, who I'd known from my previous 169 00:06:45,360 --> 00:06:49,519 work at, Philips and and doing, treatment planning 170 00:06:49,519 --> 00:06:50,500 system development. 171 00:06:51,164 --> 00:06:53,404 And he kinda came around and said, hey. 172 00:06:53,404 --> 00:06:54,604 You know, are you ready to do a 173 00:06:54,604 --> 00:06:56,044 start up with this? And we had talked 174 00:06:56,044 --> 00:06:58,524 about it several times before, and he became 175 00:06:58,524 --> 00:07:00,365 available. And I think that was sort of 176 00:07:00,365 --> 00:07:00,865 the 177 00:07:01,644 --> 00:07:02,144 entrepreneurial 178 00:07:02,524 --> 00:07:04,464 spirit injection to the whole 179 00:07:05,449 --> 00:07:06,490 program. And, 180 00:07:07,290 --> 00:07:08,970 it's just the timing was right for both 181 00:07:08,970 --> 00:07:10,490 of us, and I had a team here 182 00:07:10,490 --> 00:07:11,709 already ready to go. 183 00:07:12,089 --> 00:07:14,329 And so, yeah, we we just went ahead 184 00:07:14,329 --> 00:07:16,009 and did it. And with his knowledge of 185 00:07:16,009 --> 00:07:19,024 doing startups already and my knowledge of what, 186 00:07:19,345 --> 00:07:21,504 we wanted to achieve, it was just the 187 00:07:21,504 --> 00:07:21,985 perfect, 188 00:07:22,625 --> 00:07:24,384 perfect timing and perfect group to do it 189 00:07:24,384 --> 00:07:24,884 with. 190 00:07:25,425 --> 00:07:27,105 Okay. So so this is all led to 191 00:07:27,105 --> 00:07:28,404 sort of the the commercialization 192 00:07:28,944 --> 00:07:30,964 of this Plan AI tool. 193 00:07:31,430 --> 00:07:32,889 Now now this this platform, 194 00:07:33,349 --> 00:07:37,289 it incorporates both predictive planning and peer review. 195 00:07:37,589 --> 00:07:38,729 So first of all, 196 00:07:39,110 --> 00:07:42,069 can you describe what predictive planning means in 197 00:07:42,069 --> 00:07:43,930 the context of radiation therapy? 198 00:07:44,550 --> 00:07:47,535 Yeah. Sure. So predictive planning just, you know, 199 00:07:47,535 --> 00:07:50,435 the idea is is for a given patient, 200 00:07:50,495 --> 00:07:51,555 if I can predict 201 00:07:52,014 --> 00:07:54,014 the expected dose that I should be able 202 00:07:54,014 --> 00:07:55,235 to achieve for that patient, 203 00:07:56,095 --> 00:07:56,595 then 204 00:07:57,214 --> 00:07:58,035 that makes 205 00:07:58,574 --> 00:08:00,595 an easier problem to solve 206 00:08:00,990 --> 00:08:03,470 for when you're doing treatment planning. Treatment planning 207 00:08:03,470 --> 00:08:03,970 involves 208 00:08:05,389 --> 00:08:06,129 a a a dosimetrist 209 00:08:07,149 --> 00:08:07,649 specifying 210 00:08:08,670 --> 00:08:09,889 dosimetric objectives 211 00:08:10,350 --> 00:08:11,250 to the system, 212 00:08:11,790 --> 00:08:15,230 asking it to then optimize the delivery of 213 00:08:15,230 --> 00:08:15,730 radiation 214 00:08:16,264 --> 00:08:18,745 to meet these objectives that the dosimeter is 215 00:08:18,745 --> 00:08:19,245 specifying. 216 00:08:19,944 --> 00:08:22,745 And common practice now is nobody really knows 217 00:08:22,745 --> 00:08:25,064 what the right objectives are to to even 218 00:08:25,064 --> 00:08:27,165 put in the system. So it's a it's 219 00:08:27,464 --> 00:08:29,704 a a trial and error process trying to 220 00:08:29,704 --> 00:08:32,070 push the dose down here, and you might 221 00:08:32,070 --> 00:08:34,070 be trying to overachieve in one area and 222 00:08:34,070 --> 00:08:36,549 underachieve in other areas, and you don't really 223 00:08:36,549 --> 00:08:38,950 know it because there's so many things going 224 00:08:38,950 --> 00:08:41,929 on. And so by giving a prediction from 225 00:08:42,149 --> 00:08:44,809 from these spatial relationships in this prior data, 226 00:08:45,590 --> 00:08:47,049 it gives a nice 227 00:08:47,375 --> 00:08:49,634 rational set of objectives to that optimization, 228 00:08:50,654 --> 00:08:51,794 and that allows 229 00:08:52,975 --> 00:08:55,294 the the optimization algorithm that's already there in 230 00:08:55,294 --> 00:08:57,615 the planning systems to go towards a good 231 00:08:57,615 --> 00:09:00,274 solution and a more much more solvable problem. 232 00:09:00,620 --> 00:09:02,940 So that's the idea of predictive planning is 233 00:09:02,940 --> 00:09:03,519 is basically 234 00:09:04,059 --> 00:09:04,559 giving 235 00:09:04,940 --> 00:09:07,440 the right objective to those planning systems. 236 00:09:08,459 --> 00:09:10,240 On the peer review side, 237 00:09:10,860 --> 00:09:13,985 that's peer review is when we look, physician 238 00:09:14,144 --> 00:09:15,605 a peer physician looks 239 00:09:16,225 --> 00:09:17,985 at every treatment plan that goes through to 240 00:09:17,985 --> 00:09:19,845 evaluate it for quality and safety. 241 00:09:20,225 --> 00:09:21,365 And in that regard, 242 00:09:21,825 --> 00:09:24,065 it's very fast, and people don't really know 243 00:09:24,065 --> 00:09:25,745 how good a plan you can generate. So 244 00:09:25,745 --> 00:09:27,120 is this a good plan or not? 245 00:09:27,679 --> 00:09:29,759 Depends on the patient's anatomy. It's very hard 246 00:09:29,759 --> 00:09:31,779 to decipher that in a three d sense. 247 00:09:31,919 --> 00:09:34,100 And so by providing a predicted 248 00:09:34,799 --> 00:09:35,299 dose, 249 00:09:35,759 --> 00:09:38,580 and dose objectives and dose goals like this, 250 00:09:38,639 --> 00:09:40,865 it allows a very quick review 251 00:09:41,325 --> 00:09:42,784 to see and highlight, 252 00:09:43,085 --> 00:09:45,184 is this a high quality plan or not? 253 00:09:45,485 --> 00:09:47,725 And so that's the peer review setting. So 254 00:09:47,725 --> 00:09:49,664 it's just two different settings. One's a dosimeter 255 00:09:49,725 --> 00:09:51,485 just using it in a in in a 256 00:09:51,485 --> 00:09:54,304 planning sense. The other is a peer physician 257 00:09:54,524 --> 00:09:55,825 evaluating the quality 258 00:09:56,179 --> 00:09:57,799 of a plan that's been generated. 259 00:09:58,899 --> 00:10:01,940 Okay. So so just, basically, these objectives, they're 260 00:10:01,940 --> 00:10:04,259 to sort of maximize the dose delivered to 261 00:10:04,259 --> 00:10:06,259 the target, the tumor that you're trying to 262 00:10:06,259 --> 00:10:06,759 destroy, 263 00:10:07,379 --> 00:10:11,264 minimize any radiation to everywhere else, particularly sort 264 00:10:11,264 --> 00:10:13,205 of vital organs that might be nearby. 265 00:10:14,384 --> 00:10:15,044 And then 266 00:10:15,504 --> 00:10:15,825 the, 267 00:10:16,625 --> 00:10:17,605 the peer review 268 00:10:18,065 --> 00:10:20,304 process is a sort of a a second 269 00:10:20,304 --> 00:10:22,384 check to make sure that this has all 270 00:10:22,384 --> 00:10:22,884 worked 271 00:10:23,264 --> 00:10:23,764 okay. 272 00:10:24,360 --> 00:10:26,919 And then sort of what what types of 273 00:10:26,919 --> 00:10:28,059 anomalies can 274 00:10:28,519 --> 00:10:30,779 AI detect, for example, in 275 00:10:31,240 --> 00:10:31,980 the prescriptions, 276 00:10:32,519 --> 00:10:33,019 contours, 277 00:10:33,480 --> 00:10:35,559 treatment plans? What what sort of things does 278 00:10:35,559 --> 00:10:36,220 it find? 279 00:10:36,774 --> 00:10:39,894 Yeah. And yeah. So with, anomaly detection more 280 00:10:39,894 --> 00:10:42,375 focuses in on peer review, but also can 281 00:10:42,375 --> 00:10:45,654 be used as a preplanning evaluation step. But, 282 00:10:46,615 --> 00:10:47,975 you know, one of the peer review is 283 00:10:47,975 --> 00:10:50,578 a very hurried process. It's you're you know, 284 00:10:50,578 --> 00:10:52,736 you have a peer trying to look at 285 00:10:52,736 --> 00:10:54,893 all these plans coming through, so it's very 286 00:10:54,893 --> 00:10:57,051 fast paced. And evaluating contours can be a 287 00:10:57,051 --> 00:10:59,478 little bit of a challenge. So we've done 288 00:10:59,478 --> 00:11:01,905 some work, in the past looking at simple 289 00:11:01,905 --> 00:11:02,445 things like 290 00:11:03,004 --> 00:11:04,384 is, is a contour 291 00:11:04,845 --> 00:11:07,725 is it missing slices missing contoured slices and 292 00:11:07,725 --> 00:11:09,024 things like that? Does it, 293 00:11:09,725 --> 00:11:11,884 does it have discontinuities in it? So those 294 00:11:11,884 --> 00:11:13,985 are very basic and maybe not AI, 295 00:11:14,365 --> 00:11:16,764 but can sort of evaluate the quality of 296 00:11:16,764 --> 00:11:19,200 a contour, which is, you know, the identification 297 00:11:19,200 --> 00:11:20,959 of anatomy on a on, like, a CT 298 00:11:20,959 --> 00:11:21,459 scan. 299 00:11:21,759 --> 00:11:23,759 But other work that you can do with 300 00:11:23,759 --> 00:11:25,139 more AI is 301 00:11:25,759 --> 00:11:27,279 look at all the rectums that are in 302 00:11:27,279 --> 00:11:29,759 the system. And if your contour maybe goes 303 00:11:29,759 --> 00:11:31,139 too far into the sigmoid, 304 00:11:31,835 --> 00:11:34,475 then we can detect, hey. This rectum has 305 00:11:34,475 --> 00:11:36,315 an anomaly. It doesn't look like all the 306 00:11:36,315 --> 00:11:37,134 other rectums. 307 00:11:37,674 --> 00:11:38,815 So have you 308 00:11:39,195 --> 00:11:41,835 miscontoured that or contoured it beyond where it 309 00:11:41,835 --> 00:11:43,434 should be? And these are the kinds of 310 00:11:43,434 --> 00:11:45,595 things that physicians might look at in peer 311 00:11:45,595 --> 00:11:47,850 review, and this can then highlight and bring 312 00:11:47,850 --> 00:11:50,909 attention to things that might be anomalous 313 00:11:51,289 --> 00:11:52,829 so that they don't get overlooked. 314 00:11:53,769 --> 00:11:56,089 Prescriptions also are a a bit of a 315 00:11:56,089 --> 00:11:56,589 challenge. 316 00:11:57,450 --> 00:11:59,389 Prescriptions is as much as, 317 00:11:59,690 --> 00:12:02,184 you know, they are the physician's sort of 318 00:12:02,485 --> 00:12:04,565 realm and responsibility, but they're one of the 319 00:12:04,565 --> 00:12:07,524 hardest things to double check because a peer 320 00:12:07,524 --> 00:12:10,004 peer physician doesn't really know everything about that 321 00:12:10,004 --> 00:12:11,705 patient. What is the dose 322 00:12:12,245 --> 00:12:14,485 per fraction and how many fractions do you 323 00:12:14,485 --> 00:12:15,299 wanna deliver? 324 00:12:15,620 --> 00:12:17,940 And depending on the disease site, the location 325 00:12:17,940 --> 00:12:18,839 of the tumor, 326 00:12:19,220 --> 00:12:21,159 the diagnosis, and the intent, 327 00:12:21,620 --> 00:12:24,039 you know, whether it's curative or palliative intent, 328 00:12:24,259 --> 00:12:25,000 these prescriptions 329 00:12:26,659 --> 00:12:29,524 are kind of under protocol. And so AI 330 00:12:29,524 --> 00:12:33,125 can highlight when prescriptions are anomalies or different 331 00:12:33,125 --> 00:12:33,945 than normal, 332 00:12:34,325 --> 00:12:36,565 in which case they deserve a double check 333 00:12:36,565 --> 00:12:38,985 or a second look. There have been instances 334 00:12:39,125 --> 00:12:39,865 where somebody's 335 00:12:40,404 --> 00:12:40,904 prescribed 336 00:12:41,319 --> 00:12:44,199 six hundred centigrade per fraction for five fractions, 337 00:12:44,199 --> 00:12:46,379 but they really meant five hundred degree 338 00:12:46,839 --> 00:12:49,480 for six fractions, for example. And those kinds 339 00:12:49,480 --> 00:12:51,959 of anomalies can be detected from a safety 340 00:12:51,959 --> 00:12:52,459 perspective. 341 00:12:53,884 --> 00:12:55,404 Okay. So it's just picking out things that 342 00:12:55,404 --> 00:12:56,924 look a bit unusual, and then it can 343 00:12:56,924 --> 00:12:58,924 highlight to the the people that are working 344 00:12:58,924 --> 00:13:01,105 on the system, make sure everything's okay. 345 00:13:01,485 --> 00:13:03,184 Right. But unusual is a 346 00:13:03,804 --> 00:13:06,365 a difficult word. Right? Because unusual is it's 347 00:13:06,365 --> 00:13:08,684 a pretty broad base of inputs to determine 348 00:13:08,684 --> 00:13:09,424 what's unusual. 349 00:13:09,850 --> 00:13:10,350 Yeah. 350 00:13:11,610 --> 00:13:13,470 So, I mean, the Plan AI models, 351 00:13:14,009 --> 00:13:15,629 these are constructed 352 00:13:16,009 --> 00:13:16,830 and trained 353 00:13:17,210 --> 00:13:20,090 using the OncoSpace data lake. Now this is 354 00:13:20,090 --> 00:13:22,910 that knowledge database of all the previous treatments. 355 00:13:23,355 --> 00:13:24,634 Can you tell us a bit more about 356 00:13:24,634 --> 00:13:25,695 this data lake? 357 00:13:26,554 --> 00:13:28,955 Yeah. So in the early days, when when 358 00:13:28,955 --> 00:13:31,434 we first started, we developed a large SQL 359 00:13:31,434 --> 00:13:33,855 database. We wanted all of these 360 00:13:34,554 --> 00:13:38,014 shape relationship features and dosimetry features and everything 361 00:13:38,759 --> 00:13:41,879 directly queryable from a from a from a 362 00:13:41,879 --> 00:13:44,200 database. And we had originally designed it in 363 00:13:44,200 --> 00:13:44,700 SQL, 364 00:13:45,160 --> 00:13:47,399 and the SQL link the SQL language is, 365 00:13:47,399 --> 00:13:49,000 you know, very nice for being able to 366 00:13:49,000 --> 00:13:51,720 query out data and manage and sift through 367 00:13:51,720 --> 00:13:52,379 the data. 368 00:13:53,964 --> 00:13:54,625 And then, 369 00:13:55,725 --> 00:13:58,705 but when, the company was formed, we recognized 370 00:13:58,845 --> 00:13:59,664 that there was, 371 00:14:00,044 --> 00:14:02,684 you know, some age to that and some 372 00:14:02,684 --> 00:14:05,804 linkages to machine learning algorithms and stuff that 373 00:14:05,804 --> 00:14:07,670 weren't quite as clean as some of the 374 00:14:07,670 --> 00:14:09,850 more modern technology that was available. 375 00:14:10,790 --> 00:14:11,830 So in the, 376 00:14:12,470 --> 00:14:15,509 in the the Plan AI now data model 377 00:14:15,509 --> 00:14:18,710 or data lake, there we extracted all of 378 00:14:18,710 --> 00:14:21,450 the different shape relationship features and shape complexity 379 00:14:21,670 --> 00:14:22,125 features 380 00:14:22,924 --> 00:14:24,625 and put them into a parquet, 381 00:14:25,644 --> 00:14:27,264 database up in the cloud. 382 00:14:27,644 --> 00:14:30,304 And in that cloud now, it's linked to 383 00:14:30,445 --> 00:14:32,625 a lot of the machine learning tools available 384 00:14:32,684 --> 00:14:35,420 in the Azure platform. So it even it 385 00:14:35,420 --> 00:14:37,500 made it made a data lake much more 386 00:14:37,500 --> 00:14:39,920 amenable to applying some of the machine learning 387 00:14:40,220 --> 00:14:43,820 algorithms to it. We still maintain the, SQL 388 00:14:43,820 --> 00:14:46,080 data lake here at at Johns Hopkins 389 00:14:46,540 --> 00:14:50,000 where we primarily focus that, our efforts on 390 00:14:50,595 --> 00:14:54,115 toxicity predictions and looking at spatial patterns of 391 00:14:54,115 --> 00:14:56,355 dose and things like that. So it's taken 392 00:14:56,355 --> 00:14:58,355 a couple of different forms, but on the, 393 00:14:59,075 --> 00:15:01,014 yeah, on the plan AI side, it's, 394 00:15:01,475 --> 00:15:04,274 I would say, streamlined for the specific task 395 00:15:04,274 --> 00:15:05,129 of those prediction. 396 00:15:06,090 --> 00:15:08,730 Okay. And what does the model training process 397 00:15:08,730 --> 00:15:09,389 look like? 398 00:15:10,009 --> 00:15:10,889 Yeah. So, 399 00:15:11,450 --> 00:15:13,289 so with the the model, that was another 400 00:15:13,289 --> 00:15:14,970 thing that, you know, in our earlier work, 401 00:15:14,970 --> 00:15:15,629 we had 402 00:15:16,090 --> 00:15:18,465 sort of a simplistic or model of doing 403 00:15:18,465 --> 00:15:20,465 the dose prediction where we knew we could 404 00:15:20,465 --> 00:15:22,404 do a a pretty good job back then. 405 00:15:22,945 --> 00:15:24,384 Julie Shade, who is actually, 406 00:15:24,785 --> 00:15:28,144 the student here, data scientist that had worked 407 00:15:28,144 --> 00:15:30,370 with me in the in the past, She 408 00:15:30,370 --> 00:15:31,429 joined the company 409 00:15:31,889 --> 00:15:32,370 and, 410 00:15:33,730 --> 00:15:36,449 really took a whole another look at, the 411 00:15:36,449 --> 00:15:38,769 model and the training the model and came 412 00:15:38,769 --> 00:15:39,649 up with a nice, 413 00:15:40,129 --> 00:15:42,470 random forest ensemble model. And 414 00:15:42,929 --> 00:15:44,954 there's a lot of things. So the first 415 00:15:44,954 --> 00:15:46,475 step you have to do is really good 416 00:15:46,475 --> 00:15:47,855 data curation, and 417 00:15:48,475 --> 00:15:50,554 we had a lot more data to look 418 00:15:50,554 --> 00:15:51,774 at and a lot more, 419 00:15:53,595 --> 00:15:56,235 anatomical regions to look at. So one of 420 00:15:56,235 --> 00:15:58,174 the first things we need to do was 421 00:15:58,649 --> 00:16:01,790 curate the data by standardizing the org normal, 422 00:16:02,170 --> 00:16:03,870 a normal anatomy naming. 423 00:16:04,170 --> 00:16:06,730 So there's a task task group two, t 424 00:16:06,730 --> 00:16:09,389 g two sixty three from the APM defines 425 00:16:09,610 --> 00:16:10,110 a 426 00:16:10,649 --> 00:16:13,050 standardized structure naming model, and not all the 427 00:16:13,050 --> 00:16:14,110 data has that. 428 00:16:15,294 --> 00:16:16,995 So so she wrote some fancy, 429 00:16:18,014 --> 00:16:19,074 kind of learned 430 00:16:19,375 --> 00:16:19,875 tools 431 00:16:20,495 --> 00:16:23,375 that can, do automatic name mapping, and it 432 00:16:23,375 --> 00:16:25,154 also does automatic target, 433 00:16:26,174 --> 00:16:26,674 identifications 434 00:16:27,054 --> 00:16:29,750 because treatment plans can have multiple targets, two, 435 00:16:29,750 --> 00:16:32,070 three, and four targets within a given treatment 436 00:16:32,070 --> 00:16:32,809 plan. And 437 00:16:33,429 --> 00:16:35,690 that level of curation really helped, 438 00:16:36,549 --> 00:16:39,049 process large much larger amounts 439 00:16:39,350 --> 00:16:41,049 of, data for the model. 440 00:16:41,544 --> 00:16:43,384 And then once we had all the shape 441 00:16:43,384 --> 00:16:46,024 relationship features and shape complexity features and all 442 00:16:46,024 --> 00:16:46,684 the doses, 443 00:16:47,144 --> 00:16:49,485 we could then train the models by anatomical 444 00:16:49,625 --> 00:16:52,684 region. So we have a pelvis model that 445 00:16:52,745 --> 00:16:55,165 includes male and female, so prostate, 446 00:16:55,769 --> 00:16:56,269 GYN, 447 00:16:56,970 --> 00:16:59,690 treatments. We have a thoracic model that includes 448 00:16:59,690 --> 00:17:00,750 lung and esophagus. 449 00:17:01,129 --> 00:17:03,769 We have an abdomen model. And so for 450 00:17:03,769 --> 00:17:04,910 each of these models, 451 00:17:05,529 --> 00:17:07,869 we build or for each of these anatomical 452 00:17:07,930 --> 00:17:09,914 regions, we build a model for each 453 00:17:10,475 --> 00:17:12,815 organ at risk to predict the doses, 454 00:17:13,355 --> 00:17:14,894 based on that feature set. 455 00:17:15,434 --> 00:17:18,154 Then and that's with the random forest algorithm. 456 00:17:18,154 --> 00:17:20,795 And the the shade relationship features actually really 457 00:17:20,795 --> 00:17:22,734 do a nice job of separating out 458 00:17:23,519 --> 00:17:24,980 similar patients and separating 459 00:17:25,279 --> 00:17:26,500 them out through the trees. 460 00:17:27,119 --> 00:17:28,500 Then we just do a training. 461 00:17:28,880 --> 00:17:30,259 Do we do a five fold, 462 00:17:31,200 --> 00:17:34,080 internal validation model during the training to make 463 00:17:34,080 --> 00:17:36,400 sure that the predictions are good on the 464 00:17:36,400 --> 00:17:37,619 internal dataset. 465 00:17:38,455 --> 00:17:40,695 And then after that, we've done an external 466 00:17:40,695 --> 00:17:41,195 validation, 467 00:17:42,375 --> 00:17:42,775 at, 468 00:17:43,975 --> 00:17:44,875 multiple institutions, 469 00:17:45,335 --> 00:17:47,355 including both Johns Hopkins and Montefiore. 470 00:17:49,975 --> 00:17:51,630 I guess one other I guess one other 471 00:17:52,190 --> 00:17:53,549 quick aspect about that, 472 00:17:54,509 --> 00:17:57,869 training is is whenever we drive optimization with 473 00:17:57,869 --> 00:17:58,529 a predictive 474 00:17:59,150 --> 00:17:59,650 planning, 475 00:18:00,109 --> 00:18:01,950 we always wanna be pushing towards what we 476 00:18:01,950 --> 00:18:03,650 call a best achievable dose. 477 00:18:04,214 --> 00:18:06,054 So the prediction if you do a regular 478 00:18:06,054 --> 00:18:08,134 machine learning, you would predict sort of an 479 00:18:08,134 --> 00:18:10,714 expected dose or an average dose across patients. 480 00:18:10,934 --> 00:18:13,015 You never wanna drive a treatment plan towards 481 00:18:13,015 --> 00:18:15,255 the average dose because then every plan you 482 00:18:15,255 --> 00:18:17,654 generate will be happy being the average. We 483 00:18:17,654 --> 00:18:20,690 don't wanna do that. So we predict both 484 00:18:20,750 --> 00:18:22,589 the expected dose, which would be sort of 485 00:18:22,589 --> 00:18:23,970 like the average for patients, 486 00:18:24,509 --> 00:18:27,710 and we predict a fifth percentile level, which 487 00:18:27,710 --> 00:18:29,570 is what we would call the best achievable 488 00:18:29,710 --> 00:18:32,430 dose. So we drive plan optimization to the 489 00:18:32,430 --> 00:18:33,250 best achievable, 490 00:18:33,835 --> 00:18:36,015 hoping that we end up somewhere between 491 00:18:36,315 --> 00:18:38,394 the average and that best achievable or the 492 00:18:38,394 --> 00:18:38,894 expected. 493 00:18:39,914 --> 00:18:42,075 So our model's trained for both both of 494 00:18:42,075 --> 00:18:42,815 those predictions. 495 00:18:44,235 --> 00:18:44,735 Okay. 496 00:18:45,115 --> 00:18:47,880 So another thing to think about is that 497 00:18:47,880 --> 00:18:49,099 when you're implementing 498 00:18:49,400 --> 00:18:51,339 more any new technology in the clinic, 499 00:18:51,640 --> 00:18:53,480 it's important that it fits in with an 500 00:18:53,480 --> 00:18:55,500 existing treatment workflow workflow. 501 00:18:56,119 --> 00:18:56,619 So 502 00:18:57,079 --> 00:18:58,940 what does clinic ready mean 503 00:18:59,275 --> 00:19:00,815 for these AI tools? 504 00:19:01,674 --> 00:19:02,174 Right. 505 00:19:02,554 --> 00:19:04,394 So there's a couple of things that we 506 00:19:04,394 --> 00:19:04,894 incorporated 507 00:19:05,275 --> 00:19:05,775 in. 508 00:19:06,875 --> 00:19:07,615 One is 509 00:19:08,315 --> 00:19:09,454 we we've 510 00:19:10,154 --> 00:19:12,875 built for an anatomical region, we built models 511 00:19:12,875 --> 00:19:14,255 out of a very wide 512 00:19:14,730 --> 00:19:16,349 range of treatment protocols. 513 00:19:17,049 --> 00:19:17,869 And so 514 00:19:18,250 --> 00:19:21,049 and then radiation therapy is pretty protocol driven. 515 00:19:21,049 --> 00:19:23,690 We know what technique we're gonna use. We 516 00:19:23,690 --> 00:19:25,609 know what our clinic our dose goals are 517 00:19:25,609 --> 00:19:27,769 for different structures and things like that from 518 00:19:27,769 --> 00:19:28,714 a clinical 519 00:19:29,414 --> 00:19:31,035 perspective and how we wanna treat. 520 00:19:31,414 --> 00:19:33,194 What we don't know is the patient specific 521 00:19:33,815 --> 00:19:34,794 part of that. 522 00:19:35,255 --> 00:19:35,755 And 523 00:19:36,934 --> 00:19:37,434 so 524 00:19:38,855 --> 00:19:39,355 one 525 00:19:39,894 --> 00:19:42,720 strategy that other vendors and things have used 526 00:19:42,720 --> 00:19:44,740 is that for a given single protocol, 527 00:19:45,200 --> 00:19:47,119 we train some kind of a model to 528 00:19:47,119 --> 00:19:48,740 treat that one protocol. 529 00:19:49,359 --> 00:19:51,680 And then that means that anybody using it 530 00:19:51,680 --> 00:19:53,759 has to train for every single protocol they 531 00:19:53,759 --> 00:19:55,154 may or may not wanna use. 532 00:19:55,955 --> 00:19:58,515 And that was that's very complex and very 533 00:19:58,515 --> 00:20:00,615 hard to implement in a clinical setting. 534 00:20:00,914 --> 00:20:03,474 So our models, we have a for any 535 00:20:03,555 --> 00:20:06,134 for each anatomical region, we have a very 536 00:20:06,595 --> 00:20:09,394 large variety of different types of patients and 537 00:20:09,394 --> 00:20:11,460 treatments in the data because it's a large 538 00:20:11,460 --> 00:20:11,960 dataset, 539 00:20:12,259 --> 00:20:14,119 and the protocols fall under 540 00:20:14,420 --> 00:20:16,099 it. So we made sure that you could 541 00:20:16,099 --> 00:20:18,519 write any protocol for treatment that you wanted, 542 00:20:18,980 --> 00:20:21,059 and the same prediction model will be able 543 00:20:21,059 --> 00:20:23,234 to predict for that protocol. So that means 544 00:20:23,234 --> 00:20:24,055 any user 545 00:20:24,595 --> 00:20:26,755 can put in however they really wanna do 546 00:20:26,755 --> 00:20:28,674 things and the predictions work. They don't have 547 00:20:28,674 --> 00:20:30,275 to retrain anything. They don't have to do 548 00:20:30,275 --> 00:20:32,595 any so that's one aspect of being clinic 549 00:20:32,595 --> 00:20:34,275 ready. It's ready to go out of the 550 00:20:34,275 --> 00:20:34,680 box, 551 00:20:35,240 --> 00:20:37,160 and there's a library of protocols to start 552 00:20:37,160 --> 00:20:38,920 with, and then you can change protocols as 553 00:20:38,920 --> 00:20:40,779 you need for your own clinic. 554 00:20:41,559 --> 00:20:44,119 The other part of it being clinic ready 555 00:20:44,119 --> 00:20:45,259 is, I would say, 556 00:20:45,880 --> 00:20:47,580 aligning with the way, 557 00:20:48,680 --> 00:20:49,900 the planning currently 558 00:20:50,625 --> 00:20:51,125 happens. 559 00:20:51,585 --> 00:20:53,605 We currently use dose volume 560 00:20:54,384 --> 00:20:56,224 based objectives. So this is 561 00:20:57,265 --> 00:20:59,664 it's called a dose volume histogram, but we 562 00:20:59,904 --> 00:21:03,105 when treatment plans are performed and optimized, we're 563 00:21:03,105 --> 00:21:06,325 manipulating these little dose objectives on the points 564 00:21:06,809 --> 00:21:09,130 to drive the optimization, and that's exactly what 565 00:21:09,130 --> 00:21:09,869 we predict. 566 00:21:10,329 --> 00:21:11,069 So the 567 00:21:11,450 --> 00:21:13,690 the users are using their treatment planning system. 568 00:21:13,690 --> 00:21:16,089 We're merely just putting the objectives in there 569 00:21:16,089 --> 00:21:18,250 so they're not changing the whole paradigm of 570 00:21:18,250 --> 00:21:20,784 how they operate. They're not just running an 571 00:21:20,784 --> 00:21:22,544 automated plan and then not knowing what to 572 00:21:22,544 --> 00:21:23,524 do with it afterwards. 573 00:21:24,065 --> 00:21:26,384 They're being able to run this automation and 574 00:21:26,384 --> 00:21:28,304 be left in a position where they can 575 00:21:28,304 --> 00:21:31,044 do exactly what they normally do after that. 576 00:21:31,130 --> 00:21:32,910 So both of those, I would say, are 577 00:21:33,529 --> 00:21:35,710 the clinic ready part of it. Yeah. 578 00:21:36,250 --> 00:21:38,090 Okay. And and then can you describe some 579 00:21:38,090 --> 00:21:41,369 of the clinical validation studies that were performed 580 00:21:41,369 --> 00:21:43,724 for Plan AI and what they found? 581 00:21:44,585 --> 00:21:45,625 Yeah. So, obviously, 582 00:21:46,264 --> 00:21:49,105 we have to do external validation, and we 583 00:21:49,464 --> 00:21:51,644 you know, I guess they're not too challenging 584 00:21:51,704 --> 00:21:54,764 to do, but we looked back at 585 00:21:55,609 --> 00:21:57,710 recent treatment plans that we had 586 00:21:58,090 --> 00:22:00,570 and found clinical plans that have had the, 587 00:22:00,570 --> 00:22:01,630 you know, evaluation 588 00:22:02,009 --> 00:22:03,950 of the physicians in clinical practice. 589 00:22:04,570 --> 00:22:07,130 And we did this at both, Montfiore and, 590 00:22:07,450 --> 00:22:09,769 at Johns Hopkins, and we I don't even 591 00:22:09,769 --> 00:22:10,830 remember how many 592 00:22:11,275 --> 00:22:12,795 plans we did, but we ran a whole 593 00:22:12,795 --> 00:22:16,255 bunch of predictive plans blind from the clinical 594 00:22:16,634 --> 00:22:19,214 plans that were done and then did evaluation, 595 00:22:20,634 --> 00:22:23,674 between those predictive plans that were were run-in 596 00:22:23,674 --> 00:22:24,974 the in the clinical ones. 597 00:22:25,275 --> 00:22:25,775 And 598 00:22:28,360 --> 00:22:30,240 for almost all of the cases, the doses 599 00:22:30,240 --> 00:22:32,720 were improved. I mean, it was very obvious 600 00:22:32,720 --> 00:22:33,220 that, 601 00:22:33,839 --> 00:22:35,059 you know, there are certain 602 00:22:35,440 --> 00:22:37,059 parts of the volumes that 603 00:22:37,519 --> 00:22:39,519 the the dosimeter is doing their trial and 604 00:22:39,519 --> 00:22:41,359 error planning, not knowing how good they could 605 00:22:41,359 --> 00:22:41,859 do, 606 00:22:42,234 --> 00:22:44,255 left ignored. Right? And so we 607 00:22:45,034 --> 00:22:46,414 improved there pretty well. 608 00:22:47,674 --> 00:22:49,994 And so it's it was pretty clear that 609 00:22:49,994 --> 00:22:50,494 both 610 00:22:51,355 --> 00:22:53,214 plan quality and plan efficiency, 611 00:22:53,835 --> 00:22:55,900 were improved. The predictive plans, 612 00:22:56,599 --> 00:22:58,539 can be run with very few modifications, 613 00:22:59,640 --> 00:23:01,179 to the treatment objectives 614 00:23:01,480 --> 00:23:03,660 and very few iterations of the optimization 615 00:23:04,039 --> 00:23:05,880 where the clinical plans just take this trial 616 00:23:05,880 --> 00:23:07,720 and error process to make it take much 617 00:23:07,720 --> 00:23:08,220 longer. 618 00:23:08,865 --> 00:23:10,164 And then in one instance, 619 00:23:10,704 --> 00:23:13,605 we also had, blind clinician reviews. 620 00:23:13,984 --> 00:23:16,565 So we had three physicians blindly review, 621 00:23:17,424 --> 00:23:19,204 the because this was for the prostate, 622 00:23:19,744 --> 00:23:19,984 and, 623 00:23:21,019 --> 00:23:22,399 two out of three, 624 00:23:23,019 --> 00:23:25,659 physicians picked the predictive plan, and most of 625 00:23:25,659 --> 00:23:26,960 them all three predicted 626 00:23:27,419 --> 00:23:30,059 picked the, predicted plan over the clinical plans 627 00:23:30,059 --> 00:23:30,960 in that study. 628 00:23:33,775 --> 00:23:36,015 Okay. And then how does the Plan AI 629 00:23:36,015 --> 00:23:39,795 tool integrate with existing treatment planning systems? 630 00:23:40,575 --> 00:23:43,054 Right. Yeah. Actually, maybe I'm a little proud 631 00:23:43,054 --> 00:23:44,734 of this, but back when I was at, 632 00:23:45,295 --> 00:23:47,535 working for Philips and on the Pinnacle treatment 633 00:23:47,535 --> 00:23:49,839 planning system, one of the early things we 634 00:23:49,839 --> 00:23:50,640 did was write, 635 00:23:51,200 --> 00:23:52,980 a scripting interface for, 636 00:23:53,679 --> 00:23:56,240 access to our to control the planning system. 637 00:23:56,240 --> 00:23:58,079 And I think, we were the first to 638 00:23:58,079 --> 00:24:00,419 do that years ago, but the other vendors, 639 00:24:01,119 --> 00:24:04,694 RaySearch and, and Variant Eclipse, they've all followed 640 00:24:04,694 --> 00:24:06,315 suit and provided these 641 00:24:06,694 --> 00:24:09,515 scripting interfaces to their treatment planning systems. 642 00:24:09,974 --> 00:24:12,634 And so the way we've designed the protocols 643 00:24:13,174 --> 00:24:15,894 in the plan AI tool is it allows 644 00:24:15,894 --> 00:24:17,434 us to define prescriptions, 645 00:24:17,974 --> 00:24:20,680 the beams to be used, the different phases 646 00:24:20,680 --> 00:24:22,600 of treatment, which we might call a beam 647 00:24:22,600 --> 00:24:23,660 set or a plan, 648 00:24:24,120 --> 00:24:26,759 and then all of the objectives and as 649 00:24:26,759 --> 00:24:29,320 well as the clinical goals. And we define 650 00:24:29,320 --> 00:24:30,380 those in a protocol. 651 00:24:31,160 --> 00:24:33,340 And then for both of the two 652 00:24:33,720 --> 00:24:35,384 to their treatment planning vendors, 653 00:24:35,865 --> 00:24:38,345 we through that scripting interface, each element of 654 00:24:38,345 --> 00:24:41,005 that protocol can be executed through a script. 655 00:24:41,464 --> 00:24:41,964 So 656 00:24:42,585 --> 00:24:45,164 so from the planning system that the dosimetrists 657 00:24:45,384 --> 00:24:46,045 are using, 658 00:24:46,424 --> 00:24:47,644 they can request 659 00:24:48,529 --> 00:24:50,929 this protocol to be executed in their in 660 00:24:50,929 --> 00:24:51,669 their system, 661 00:24:51,970 --> 00:24:54,210 and it will go through. It'll create the 662 00:24:54,210 --> 00:24:56,470 plan. It creates each of the beam sets, 663 00:24:56,609 --> 00:24:58,869 creates all the beams, sets the prescriptions, 664 00:24:59,490 --> 00:25:01,589 and then it will load all the objectives 665 00:25:01,890 --> 00:25:04,025 into the system and have it ready for 666 00:25:04,025 --> 00:25:04,924 them to optimize 667 00:25:05,304 --> 00:25:05,964 like that. 668 00:25:06,345 --> 00:25:09,065 And, yeah, it works out quite well. And 669 00:25:09,065 --> 00:25:11,164 we were, you know, smart not to make 670 00:25:11,625 --> 00:25:14,345 commands that are too challenging. They're very they're 671 00:25:14,345 --> 00:25:15,804 very simple set of commands 672 00:25:16,319 --> 00:25:18,799 that, that the planning system vendors provide, 673 00:25:19,119 --> 00:25:20,500 open access to. 674 00:25:22,079 --> 00:25:23,220 Okay. And then 675 00:25:23,759 --> 00:25:25,839 my next question, I guess we've touched on 676 00:25:25,839 --> 00:25:27,519 this a little with the idea of being 677 00:25:27,519 --> 00:25:28,640 clinic ready, but, 678 00:25:29,394 --> 00:25:31,654 did you encounter any particular challenges 679 00:25:32,275 --> 00:25:35,315 bringing big data and AI into the clinical 680 00:25:35,315 --> 00:25:36,934 radiation oncology setting? 681 00:25:37,795 --> 00:25:39,095 Yeah. I, that's 682 00:25:39,634 --> 00:25:40,134 twofold. 683 00:25:40,674 --> 00:25:42,214 You know, interestingly enough, 684 00:25:43,075 --> 00:25:45,940 the it's not the challenges aren't really technical. 685 00:25:46,080 --> 00:25:46,900 The challenges 686 00:25:47,200 --> 00:25:49,619 are are more human related. Right? 687 00:25:50,720 --> 00:25:52,420 I guess probably one of 688 00:25:52,880 --> 00:25:53,859 the more systemic 689 00:25:54,240 --> 00:25:54,740 challenges 690 00:25:55,039 --> 00:25:56,960 is on, you know, the use of of 691 00:25:56,960 --> 00:25:59,474 medical data for training and things like that. 692 00:26:00,355 --> 00:26:01,174 There's always, 693 00:26:02,355 --> 00:26:03,174 data security. 694 00:26:04,275 --> 00:26:07,154 The IRBs are always, you know, questionable about 695 00:26:07,154 --> 00:26:09,714 how you're using data. And it's another nice 696 00:26:09,714 --> 00:26:11,714 thing about our system is the feature set 697 00:26:11,714 --> 00:26:14,115 that we need to generate from the treatment 698 00:26:14,115 --> 00:26:17,160 plans are really just these, like, mathematical features 699 00:26:17,160 --> 00:26:19,480 of shape relationships and don't involve a lot 700 00:26:19,480 --> 00:26:20,859 of identifiable information. 701 00:26:21,400 --> 00:26:23,480 But you have to get through that whenever 702 00:26:23,480 --> 00:26:25,720 you're using any kind of medical data with, 703 00:26:25,720 --> 00:26:27,704 you know, review boards and things like that. 704 00:26:29,464 --> 00:26:30,024 On the, 705 00:26:30,984 --> 00:26:32,044 I guess, implementation 706 00:26:32,424 --> 00:26:34,284 side from a clinical side, 707 00:26:36,065 --> 00:26:37,565 you know, there's a healthy skepticism 708 00:26:38,024 --> 00:26:39,164 of AI performance. 709 00:26:40,105 --> 00:26:42,299 In our in radiation therapy, there's been a 710 00:26:42,299 --> 00:26:44,319 lot of AI on auto segmentation. 711 00:26:45,019 --> 00:26:47,339 So contouring tools from images and things like 712 00:26:47,339 --> 00:26:49,900 that. And, you know, early efforts were not 713 00:26:49,900 --> 00:26:51,579 so good or not as good, and people 714 00:26:51,579 --> 00:26:53,654 had to modify things. So there's there's always 715 00:26:53,654 --> 00:26:55,195 this good healthy skepticism. 716 00:26:55,815 --> 00:26:57,975 But I think once you show to them 717 00:26:57,975 --> 00:27:00,535 that, it works and it works well and 718 00:27:00,535 --> 00:27:02,634 they still have control over it after it, 719 00:27:03,735 --> 00:27:04,875 that can be overcome. 720 00:27:05,735 --> 00:27:07,335 I have seen and and I I would 721 00:27:07,335 --> 00:27:08,779 say probably with some 722 00:27:11,000 --> 00:27:14,059 of the older planners and dosimetres, there is 723 00:27:14,279 --> 00:27:16,440 some fear of job security. People are always 724 00:27:16,440 --> 00:27:17,740 a little bit worried about 725 00:27:18,360 --> 00:27:20,220 AI taking over what they do. 726 00:27:20,680 --> 00:27:22,365 There's an art. Right? I mean, 727 00:27:22,924 --> 00:27:25,644 treatment planning and radiation therapy is you know, 728 00:27:25,644 --> 00:27:29,424 we're they're medical professionals. They're they're they're designing 729 00:27:29,484 --> 00:27:31,585 this treatment plan to care for a patient, 730 00:27:31,804 --> 00:27:33,105 and there's a lot of pride 731 00:27:33,404 --> 00:27:35,825 and art in that. And if you automate 732 00:27:35,884 --> 00:27:36,384 that, 733 00:27:37,164 --> 00:27:37,559 that's 734 00:27:38,600 --> 00:27:40,440 taking away some of that art and some 735 00:27:40,440 --> 00:27:41,340 of that pride. 736 00:27:41,880 --> 00:27:42,380 And 737 00:27:43,160 --> 00:27:43,559 it's, 738 00:27:44,519 --> 00:27:46,200 I think it's a little bit challenged to 739 00:27:46,200 --> 00:27:48,759 overcome. And and people say, you know, your 740 00:27:48,759 --> 00:27:50,600 job's not gonna be taken over by AI. 741 00:27:50,600 --> 00:27:52,200 That's gonna be taken over by the people 742 00:27:52,200 --> 00:27:53,535 that embrace AI and use 743 00:27:54,335 --> 00:27:56,095 it. Right? And and that's that's true to 744 00:27:56,095 --> 00:27:58,654 some extent. And and I guess there's other 745 00:27:58,654 --> 00:27:59,154 things. 746 00:28:00,494 --> 00:28:02,015 Yeah. I mean, I think that's a little 747 00:28:02,015 --> 00:28:02,835 bit to overcome. 748 00:28:04,734 --> 00:28:06,815 But maybe, you know, if you're if you're 749 00:28:06,815 --> 00:28:07,955 automating the monotony 750 00:28:08,569 --> 00:28:10,509 and give and and and show them that, 751 00:28:10,650 --> 00:28:12,329 really, no. We're helping you get to this 752 00:28:12,329 --> 00:28:14,829 point where where you can apply your art 753 00:28:15,529 --> 00:28:16,349 more often 754 00:28:16,650 --> 00:28:18,109 and and and more meaningfully 755 00:28:18,490 --> 00:28:21,450 rather than spending all of your practice trial 756 00:28:21,450 --> 00:28:23,325 and error trying to get to something that 757 00:28:23,804 --> 00:28:25,085 may or may not be able to be 758 00:28:25,085 --> 00:28:27,005 achieved. Well, hey. No. We'll put you right 759 00:28:27,005 --> 00:28:29,644 there where you can achieve it. Now you 760 00:28:29,644 --> 00:28:31,565 can advance it from that point forward, and 761 00:28:31,565 --> 00:28:33,744 I think that can help a lot. Yeah. 762 00:28:35,164 --> 00:28:37,024 Yeah. I guess so. It's it's, 763 00:28:37,325 --> 00:28:37,825 working 764 00:28:38,500 --> 00:28:41,059 together or using AI as a tool rather 765 00:28:41,059 --> 00:28:43,140 than replacing what you do, a tool to 766 00:28:43,140 --> 00:28:43,799 help you 767 00:28:44,179 --> 00:28:46,019 in what you're doing. So Yeah. And you 768 00:28:46,019 --> 00:28:47,940 don't want a paradigm shift. You don't wanna 769 00:28:47,940 --> 00:28:50,740 make it this totally different paradigm on how 770 00:28:50,740 --> 00:28:53,115 something gets done. Wanna, you know, give them 771 00:28:53,275 --> 00:28:54,875 put them put them in the same place 772 00:28:54,875 --> 00:28:57,295 that they're already working would help them. Yep. 773 00:28:57,515 --> 00:28:59,595 Yeah. So so looking ahead, where do you 774 00:28:59,595 --> 00:29:02,015 see predictive modeling in AI in oncology, 775 00:29:02,475 --> 00:29:02,975 say, 776 00:29:03,434 --> 00:29:05,410 five years from now, and how could it 777 00:29:05,410 --> 00:29:07,029 impact patient outcomes? 778 00:29:09,490 --> 00:29:11,650 Yeah. So, you know, right now, I think 779 00:29:11,650 --> 00:29:14,309 there's still some more advancement to go, 780 00:29:15,089 --> 00:29:18,845 more body sites and, you know, broader ranges 781 00:29:18,845 --> 00:29:19,585 of of, 782 00:29:20,044 --> 00:29:22,524 of data to improve some of the models 783 00:29:22,524 --> 00:29:23,505 to get more 784 00:29:23,884 --> 00:29:25,565 specific. So I think one of the areas 785 00:29:25,565 --> 00:29:29,265 is growing the current models to the next 786 00:29:29,644 --> 00:29:31,244 you know, to to fill out the rest 787 00:29:31,244 --> 00:29:33,059 of the body sites, things like that. 788 00:29:34,420 --> 00:29:36,700 Another thing that I always think about is, 789 00:29:36,700 --> 00:29:39,400 you know, once you get going really with 790 00:29:39,779 --> 00:29:40,600 the concept 791 00:29:41,220 --> 00:29:43,620 is getting multiple centers in sort of a 792 00:29:43,620 --> 00:29:44,120 consortium 793 00:29:45,220 --> 00:29:47,539 adding to this pool of of knowledge and 794 00:29:47,539 --> 00:29:49,154 data so that it truly becomes 795 00:29:49,775 --> 00:29:51,694 sort of the prior visions of a learning 796 00:29:51,694 --> 00:29:52,434 health system. 797 00:29:52,894 --> 00:29:54,015 You know, right now, there's been a lot 798 00:29:54,015 --> 00:29:55,775 of data collected. We built a model. You 799 00:29:55,775 --> 00:29:56,974 can train the model, and you can do 800 00:29:56,974 --> 00:29:57,634 the predictions. 801 00:29:57,934 --> 00:30:00,414 But, really, you want that data to advance 802 00:30:00,414 --> 00:30:02,138 and learn. So you so having multiple centers 803 00:30:02,138 --> 00:30:02,164 contribute back to that knowledge base as they 804 00:30:02,164 --> 00:30:02,924 advance or as they change, 805 00:30:04,230 --> 00:30:04,730 can 806 00:30:10,492 --> 00:30:10,992 be 807 00:30:16,914 --> 00:30:18,515 can be of a huge value, and and 808 00:30:18,515 --> 00:30:21,154 a community environment like that should be able 809 00:30:21,154 --> 00:30:23,075 to work. I see no reason why it 810 00:30:23,075 --> 00:30:25,015 can't work, and it it should be great. 811 00:30:26,434 --> 00:30:28,934 In terms of sort of patient outcomes 812 00:30:30,609 --> 00:30:32,130 and things like that, we do a lot 813 00:30:32,130 --> 00:30:33,029 of work here. 814 00:30:34,210 --> 00:30:36,049 One of the things with normal anatomy is 815 00:30:36,049 --> 00:30:37,589 we we tend to look at 816 00:30:38,049 --> 00:30:40,769 where the dose goes in this ball of 817 00:30:40,769 --> 00:30:41,269 anatomy, 818 00:30:41,904 --> 00:30:43,605 and it assumes that every 819 00:30:43,904 --> 00:30:47,025 location in that in that anatomy is equally 820 00:30:47,025 --> 00:30:48,005 sensitive to radiation 821 00:30:48,865 --> 00:30:49,365 and 822 00:30:49,664 --> 00:30:52,244 equally important to the function of that anatomy. 823 00:30:52,944 --> 00:30:55,265 And that's sort of because it's complex. So 824 00:30:55,265 --> 00:30:56,865 we we we try to make it simple 825 00:30:56,865 --> 00:30:58,619 and think about it that way. And 826 00:31:00,599 --> 00:31:01,960 the a lot of the work we've done 827 00:31:01,960 --> 00:31:04,279 here is to look at how a spatial 828 00:31:04,279 --> 00:31:06,599 pattern of dose on a complex piece of 829 00:31:06,599 --> 00:31:07,099 anatomy 830 00:31:07,799 --> 00:31:09,734 impacts toxicities and outcomes. 831 00:31:10,214 --> 00:31:11,575 So if, you know, if I have a 832 00:31:11,575 --> 00:31:13,355 parotid gland and I'm irradiating 833 00:31:13,974 --> 00:31:15,835 stem cell region too heavily, 834 00:31:16,214 --> 00:31:19,095 maybe I destroy the ability for recovery down 835 00:31:19,095 --> 00:31:21,115 the road. But if I'm just irradiating 836 00:31:21,414 --> 00:31:23,275 sort of the acinar cells responsible 837 00:31:23,779 --> 00:31:25,539 you know, the outer acinar cells that produce 838 00:31:25,539 --> 00:31:26,039 saliva, 839 00:31:26,659 --> 00:31:28,659 then maybe it's not so bad because it 840 00:31:28,659 --> 00:31:31,139 can recover in a year post treatment. And 841 00:31:31,139 --> 00:31:33,059 so looking at how that pattern of dose 842 00:31:33,059 --> 00:31:33,559 impacts, 843 00:31:34,179 --> 00:31:36,839 the toxicities is something that's coming. 844 00:31:38,184 --> 00:31:40,285 I think that I would call that discovery 845 00:31:40,345 --> 00:31:43,144 mode and trying to use AI to advance 846 00:31:43,144 --> 00:31:45,785 our own knowledge of how radiation impacts the 847 00:31:45,785 --> 00:31:46,285 anatomy. 848 00:31:46,744 --> 00:31:48,664 But I do see down the road that, 849 00:31:48,984 --> 00:31:51,480 some of these same concepts where we predict 850 00:31:51,480 --> 00:31:54,140 those, we could predict normal tissue outcomes 851 00:31:54,599 --> 00:31:56,359 and things like that and feed that into 852 00:31:56,359 --> 00:31:58,220 the planning process to help reduce 853 00:31:58,519 --> 00:31:59,820 that at the patient level. 854 00:32:00,839 --> 00:32:03,080 So those those, I think, are, 855 00:32:03,544 --> 00:32:04,845 some important advancements. 856 00:32:05,384 --> 00:32:05,884 Yeah. 857 00:32:06,744 --> 00:32:08,204 And do you have any advice 858 00:32:08,505 --> 00:32:12,025 for clinics interested in adopting this AI driven 859 00:32:12,025 --> 00:32:12,525 planning? 860 00:32:13,544 --> 00:32:13,944 Yeah. 861 00:32:14,585 --> 00:32:15,544 Yeah. In general, 862 00:32:15,944 --> 00:32:18,419 you know, introduce it as an assistant. Don't 863 00:32:18,419 --> 00:32:19,879 introduce it as a solution, 864 00:32:21,379 --> 00:32:23,059 you know, to get you want the the 865 00:32:23,059 --> 00:32:24,839 current people that know what they're doing 866 00:32:25,220 --> 00:32:27,619 to be able to use their knowledge even 867 00:32:27,619 --> 00:32:28,839 better and more efficiently. 868 00:32:29,940 --> 00:32:30,519 And then, 869 00:32:30,980 --> 00:32:32,875 you know, we wanna make their jobs easier 870 00:32:33,335 --> 00:32:35,255 and show them that it also improves that 871 00:32:35,255 --> 00:32:35,755 quality. 872 00:32:37,815 --> 00:32:39,974 And I think once once that overall value 873 00:32:39,974 --> 00:32:41,755 is realized and sort of 874 00:32:42,214 --> 00:32:44,294 in the clinical environment where they they see 875 00:32:44,294 --> 00:32:46,409 that, wow, you know, I can go right 876 00:32:46,409 --> 00:32:48,730 at this predicted solution. And in peer review, 877 00:32:48,730 --> 00:32:50,250 we can see, hey. This plan is at 878 00:32:50,250 --> 00:32:52,409 the level of quality that we expect, that 879 00:32:52,409 --> 00:32:54,970 we predict. We're no longer guessing in our 880 00:32:54,970 --> 00:32:57,069 minds. We're not arguing back and forth. 881 00:32:57,450 --> 00:32:58,835 Like, a lot of times with with the 882 00:32:58,835 --> 00:33:00,595 dosimeters, you know, they're they've got a plan. 883 00:33:00,595 --> 00:33:02,674 They've worked really hard getting this dose down, 884 00:33:02,674 --> 00:33:03,494 and the physicians 885 00:33:03,795 --> 00:33:05,555 looks at it and goes, oh, certainly, you 886 00:33:05,555 --> 00:33:08,115 could do better, not knowing what they've done. 887 00:33:08,115 --> 00:33:10,115 Right? And so here, it just gives them 888 00:33:10,115 --> 00:33:12,990 that that confidence that they're right there. And 889 00:33:13,309 --> 00:33:14,750 so once they have that, 890 00:33:16,269 --> 00:33:18,269 you know, once they realize that and they 891 00:33:18,269 --> 00:33:20,349 kinda remove those initial doubts and start using 892 00:33:20,349 --> 00:33:20,849 it, 893 00:33:21,470 --> 00:33:21,970 then 894 00:33:22,429 --> 00:33:24,670 you really can start bringing more value from 895 00:33:24,670 --> 00:33:24,910 it. 896 00:33:25,484 --> 00:33:27,904 You could start using it for adaptive planning 897 00:33:28,045 --> 00:33:29,505 and and other things 898 00:33:29,964 --> 00:33:30,464 because 899 00:33:30,765 --> 00:33:33,164 you've gained that level of constant confidence and 900 00:33:33,164 --> 00:33:35,404 you're not spinning your wheels on it. And 901 00:33:35,404 --> 00:33:37,345 and then you can use it to leapfrog 902 00:33:37,404 --> 00:33:39,325 some of the other technologies and buy people 903 00:33:39,325 --> 00:33:39,690 time. 904 00:33:40,169 --> 00:33:41,450 I guess one of the quotes I would 905 00:33:41,450 --> 00:33:43,690 say that I've I've often said is, you 906 00:33:43,690 --> 00:33:46,169 know, why don't we automate the easy easier 907 00:33:46,169 --> 00:33:48,829 redundant things that you can spend your quality 908 00:33:48,889 --> 00:33:51,789 time on the more difficult and challenging cases? 909 00:33:52,335 --> 00:33:53,394 Because that's where 910 00:33:53,934 --> 00:33:56,355 your your talent might be even more needed 911 00:33:56,494 --> 00:33:58,914 rather than doing the redundant over and over 912 00:33:58,974 --> 00:34:00,914 repeat things that maybe are a lot easier 913 00:34:01,375 --> 00:34:02,914 easier more easily done. 914 00:34:04,299 --> 00:34:06,700 Okay. So what what's been the most rewarding 915 00:34:06,700 --> 00:34:08,719 part of this journey for you personally? 916 00:34:10,539 --> 00:34:11,039 Yeah. 917 00:34:11,820 --> 00:34:13,820 I mean, on the predictive planning side I 918 00:34:13,820 --> 00:34:15,179 mean, on the research side, I love the 919 00:34:15,179 --> 00:34:16,860 big data and all the all the knowledge 920 00:34:16,860 --> 00:34:18,460 and discovery we can try to do with 921 00:34:18,460 --> 00:34:20,215 it. But some of that, yeah, it's still 922 00:34:20,215 --> 00:34:22,635 a little far fetched. But from the predictive 923 00:34:22,775 --> 00:34:23,755 planning side, 924 00:34:24,614 --> 00:34:26,234 you know, having my prior experience, 925 00:34:27,335 --> 00:34:28,954 building treatment planning systems 926 00:34:29,335 --> 00:34:31,414 at Phillips and ADAC, you know, we built 927 00:34:31,414 --> 00:34:35,034 a this intensity modulated radiotherapy planning system. And 928 00:34:35,710 --> 00:34:38,110 back then, you know, people were worried. Oh, 929 00:34:38,110 --> 00:34:40,190 the the system is just gonna optimize the 930 00:34:40,190 --> 00:34:41,809 delivery of this treatment plan, 931 00:34:42,190 --> 00:34:43,869 and and and our jobs are gonna go 932 00:34:43,869 --> 00:34:45,309 away and whatever. And, 933 00:34:46,190 --> 00:34:48,110 but the biggest problem we always had is 934 00:34:48,110 --> 00:34:50,984 nobody knew what the objective was. Nobody knew 935 00:34:50,984 --> 00:34:53,065 how to tell the system, this is the 936 00:34:53,065 --> 00:34:54,605 dose I expect to receive. 937 00:34:55,784 --> 00:34:57,864 Now go optimize to get it for me 938 00:34:57,864 --> 00:34:58,364 because 939 00:34:58,904 --> 00:34:59,404 you 940 00:34:59,944 --> 00:35:02,045 you couldn't know what you could do. 941 00:35:02,424 --> 00:35:04,684 And so for any given patient, you could 942 00:35:04,769 --> 00:35:06,789 ask for too much, ask for too little, 943 00:35:07,010 --> 00:35:08,690 and you get the results. So we never 944 00:35:08,690 --> 00:35:10,710 actually knew that. And for the first time, 945 00:35:11,730 --> 00:35:14,369 I I argue we actually know what our 946 00:35:14,369 --> 00:35:16,469 objective is in our treatment planning. 947 00:35:16,864 --> 00:35:17,924 And bringing that, 948 00:35:18,464 --> 00:35:20,704 you know, forward and being able to provide 949 00:35:20,704 --> 00:35:22,545 that to the rest of the world as 950 00:35:22,545 --> 00:35:25,025 a as a tool to use in clinics, 951 00:35:25,025 --> 00:35:26,085 it affects everybody. 952 00:35:26,704 --> 00:35:26,944 And, 953 00:35:27,664 --> 00:35:29,184 so now I can say that. I kind 954 00:35:29,184 --> 00:35:32,050 of rounded out that that intensity mod you 955 00:35:32,050 --> 00:35:34,070 know, that optimization, that plan optimization. 956 00:35:34,449 --> 00:35:36,710 We've rounded it out and finally told people 957 00:35:37,010 --> 00:35:39,010 what they can achieve. Now we just have 958 00:35:39,010 --> 00:35:40,550 to get everybody using it. 959 00:35:41,170 --> 00:35:41,670 Excellent. 960 00:35:43,055 --> 00:35:45,454 So if you could leave our listeners with 961 00:35:45,454 --> 00:35:46,595 one key takeaway, 962 00:35:47,135 --> 00:35:48,195 what would it be? 963 00:35:50,575 --> 00:35:53,055 You know, AI is all the AI tools. 964 00:35:53,055 --> 00:35:53,695 Right? It's, 965 00:35:54,735 --> 00:35:57,519 for the most part, it's all trained on 966 00:35:57,519 --> 00:36:00,179 historical data, prior data. Right? So, really, 967 00:36:01,199 --> 00:36:02,960 it's kind of just telling you what you 968 00:36:02,960 --> 00:36:04,960 can do from what you've been able to 969 00:36:04,960 --> 00:36:05,859 do in the past. 970 00:36:06,239 --> 00:36:08,315 And if you can remember, you're just you're 971 00:36:08,315 --> 00:36:11,055 just have a sophisticated system of of recalling 972 00:36:11,115 --> 00:36:12,175 what you've done before, 973 00:36:12,474 --> 00:36:14,235 showing you what you can do now on 974 00:36:14,235 --> 00:36:16,175 a on an individual patient basis. 975 00:36:16,474 --> 00:36:18,635 And bringing that to light in a useful 976 00:36:18,635 --> 00:36:19,135 way 977 00:36:19,675 --> 00:36:21,994 is just it's helpful. It's and and people 978 00:36:21,994 --> 00:36:22,735 need to 979 00:36:23,250 --> 00:36:25,889 see that and believe that and and appreciate 980 00:36:25,889 --> 00:36:26,710 that. And, 981 00:36:27,730 --> 00:36:29,030 you know, I think it also 982 00:36:29,489 --> 00:36:31,750 in a hurried clinic, in a busy clinic, 983 00:36:31,809 --> 00:36:34,710 you have dosimetres with different levels of experience, 984 00:36:35,424 --> 00:36:37,045 and it really does kind of 985 00:36:37,424 --> 00:36:39,284 level the playing field or bring 986 00:36:39,664 --> 00:36:41,284 a more uniform quality 987 00:36:41,984 --> 00:36:42,484 across 988 00:36:43,025 --> 00:36:45,824 whatever environment you're in, whether, you know, you're 989 00:36:45,824 --> 00:36:47,984 in different countries around the world or different 990 00:36:47,984 --> 00:36:51,160 environments or different busyness of the clinics or 991 00:36:51,160 --> 00:36:53,880 different, you know, experience level of the treatment 992 00:36:53,880 --> 00:36:54,380 planner. 993 00:36:54,920 --> 00:36:56,140 This tool really 994 00:36:56,519 --> 00:36:59,239 brings all that to a uniform state, and 995 00:36:59,239 --> 00:37:02,074 gives the best quality every everywhere you go. 996 00:37:03,335 --> 00:37:04,554 Excellent. Well, 997 00:37:04,855 --> 00:37:07,514 thank you, Todd, for speaking with us today. 998 00:37:07,815 --> 00:37:09,094 Thank you for having me. It's, 999 00:37:09,894 --> 00:37:11,574 it's hard to talk about all this stuff, 1000 00:37:11,574 --> 00:37:13,230 but it's great. Thank you so much. 1001 00:37:20,109 --> 00:37:23,730 That was Todd McNutt of Johns Hopkins University 1002 00:37:24,190 --> 00:37:26,210 and the founder of OncoSpace. 1003 00:37:27,485 --> 00:37:30,465 He was speaking to Physics World's Tammy Freeman. 1004 00:37:31,085 --> 00:37:33,105 Thanks to both of them for a fascinating 1005 00:37:33,405 --> 00:37:33,905 conversation. 1006 00:37:34,605 --> 00:37:37,025 And a special thanks to Sun nuclear 1007 00:37:37,485 --> 00:37:39,184 for sponsoring this episode. 1008 00:37:39,789 --> 00:37:42,989 I'm Hamish Johnston, and our producer is Fred 1009 00:37:42,989 --> 00:37:43,489 Iles. 1010 00:37:44,029 --> 00:37:45,889 We'll be back again next week.