00:00:02 - 00:01:18
Aleks: Good morning. Welcome my digital pathology trailblazers. How are you today? Welcome to the Digipath Digest number 27. Can you imagine? We have done already 27 of these episodes and several of you have been here several times. So, whenever you join the live stream, let me know in the comments where you are tuning in from. What time is it for you? It's 6:00 a.m. and for me 6:03. I'm actually 2 minutes late. Sorry for that. And let me know that you hear me well and just like give me a thumbs up and that everything
00:00:41 - 00:02:09
is working. Sometimes the LinkedIn comments uh come through a little late. So don't get discouraged by me repeating it several times. And but thank you so much for joining and let's start. Today it's going to be kind of short and sweet because we only have four abstracts that I wanted to go through. So let's start with the first one which is iminoista chemistry guided segmentation of B9 epithelial cells in situations and invasive epithelial cells in breast cancer slides. Let me make myself small.
00:01:23 - 00:03:02
Okay. So, is my pen working? Yes, it is. Amazing. So, um it's in plus one and the authors are from Tronheim, Norway. And I liked this uh abstract because they they did this imuninohistochemistry guided segmentation. And I always give this example when um they were doing the chameleon challenge. Uh there there is this chameleon challenge. Uh it was to detect breast cancer metastasis or like metastasis cancer metastasis in lymph nodes and uh their ground truth. It was kind of a double double ground truth
00:02:14 - 00:03:16
because they had uh iminoistochemistry and I see new people joining. Thank you so much. So in the chameleon challenge they had the imuno hystochemistry to cytoin to stain the epithelial cells and they also had pathologist annotate that and that's kind of a similar uh framework in this paper. If you are just joining let me know in the comments where you're tuning in from. So let's see what they exactly did and how it worked. So um obviously at the beginning we know that digital pathology
00:02:45 - 00:03:59
is fantastic right and it enables segmentation with um AI AI based image analysis and uh it could improve diagnostic efficiency and find associations between morphological features and clinical outcomes. That's our holy grail and I see more people joining. Thank you so much. Thank you so much. let me know in the comments um where you're tuning in from and just like give me a hi in the comments so that I can see that you guys are really there and that you can see me and hear me well. So um they did develop a model like this
00:03:22 - 00:04:43
segmentation model for breast cancer and they wanted to identify invasive epithelial cells and separate these from B9 epithelial cells and also in situations. uh that was that was the importance it was important uh to have all the epithelial cells and then divide them into subcategories. So they had this unit for segmentation of epithelial cells in H& uh and they generated epithelial ground truths by imunoistochemistry. So they restained H& sections with cytoin. So they had the same sections
00:04:02 - 00:05:09
which which is uh super cool because uh then you don't have to like guess is it the same cell or not. Uh you know you have a little bit of ar architectural distortion of the slide when you stain and restain. It doesn't like exactly look the same because of the uh process it goes through but it's close enough better even than serial sections like better even better than serial sections. and then combined it with pathologist annotations and and they had tissue microarrays. So this is important. Um
00:04:36 - 00:06:00
tissue microarrays are these like um you know these slides with little circles on it and these are the you like there are special machines where you can cut out of um histoathology blocks out of paraffhin blocks these holes and these are tissue microarrays. Let me get rid of this. [Music] So they had these from 839 patients and they also had horse light images from two patients. It's so so cool that whole light images were just from two patients. Um okay and uh they used to train this model, right? Uh so the
00:05:20 - 00:06:43
sections that uh sections were derived from four breast cancer cohorts. So that's important because they didn't just do like one cohort. Uh they had different ones uh to ensure generalizability and um there were tissue microarrays from the fifth cohort of 21 patients and this was used as a second test set. Um and they had dice scores of over 0.7 one even 0.79. um they achieved these four invasive epithelial cells, B9 epithelial cells and insecto lesions respectively. And then what else they did? Let me like get
00:06:02 - 00:07:36
rid of this a little bit because it's too much too much you cannot see. Um what uh in addition what they did in addition and they also used this qualitative scoring by the pathologist. So this scoring was 0 to5 and the best results were reached for epithelial and invasive epithelium uh with scores 4.7 and 4.4 respectively. Um and benign epithelium and insecto lesions were a little less 3.7 and two for be for incal lesions. So they were not happy with that. So they proposed this proposed model segmented
00:06:48 - 00:08:09
epithelial cells well but further work is needed for this subclassification into benign in C2 and invasive cells. Am I surprised? Not really. I'm not surprised by that because if it's like challenging then it's going to be challenging for the AI as well. H and I see more and more people joining. So, leave me a comment in the chat that you're here. Uh, and where you're tuning in from. Okay, another topic I kind of well, not kind of, I feel strongly about the but I feel strongly about Oh, another person
00:07:29 - 00:08:55
joining. Thank you so much. Thank you so much for coming and let me know in the chat uh where you're tuning in from. So um a topic I feel strongly about and you've heard about this if you've listened to me um is that when we have to like quantify visually stuff that a computer could quantify a little better. And one of these things that need need to be quantified for different cancers um is that tumor infiltrating lymphosytes. Um, let me move my camera a little bit. Two more infiltrating lymphosytes.
00:08:11 - 00:09:45
[Music] Um, and these tumor infiltrating lymphocytes this time were checked or or were investigated in esophagial cancer. So this is a group from China published in cancer medicine and what they did was uh they the these tills were uh have been proven to be independent important sorry important prognostic factors for various tumors but it has not been checked for esopagio squamos cell carcinoma yet. ESCC was not checked. So this is what they wanted to check if this going to be a good prognostic
00:09:03 - 00:10:40
marker. So they enrolled retrospectively 626 pathologically confirmed um esophagial cancer patients from two research centers and um they use their host images and clinical information and they developed a deep learning method to segment these lymphosytes. Um so where did they segment? So here we are in the in the area in the realm of immune oncology of uh our immune cells immune system trying to protect us defend us from cancer and um excuse me no I thought I'm going to sneeze but it's going to happen sooner or later.
00:09:52 - 00:10:55
And so that's what they wanted to check if these tils tumor infiltrating lymphocytes are going to help us um are going to be a good prognostic uh marker and they checked them in the tumor margin. They they found they um designed a model to identify tumor margin. Why tumor margin? because it's important where these tills uh immune cells go into the tumor and then based on the distance from the tumor margin uh you decide whether it's a good marker or a bad marker or you decide your thresholds and
00:10:26 - 00:11:55
distances right so it's a little bit of like spatial biology uh imunoncology this kind of stuff so they divided tissue into intraumor perumemeral and strummal region based on the distance from tumor margins So, let me just show you what they do. Like my my little drawing is going to be okay. I don't know. Let's say this is the esophagus. This is going to be the tumor. And can I change my color? Yes. This is going to be the invasive margin like exactly next to the tumor. And then you're going to have intraatumeral
00:11:10 - 00:12:37
and perittumal. And there is going to be a distance within this invasive margin. Let me know if this makes sense to you. Uh this is the tumor, right? This is the tumor. Let me know in the comments if you understand this image. I have a better one in my like resources. This is our tumor. Anyway, let me know if that was good enough. Uh so they counted teals in each of these regions and then they like did their statistics and um found different cutoff values and um then they also mitigated the selection bias with propensity score
00:11:54 - 00:13:18
uh matching analysis. Um and what were the results? They classified patients based on the cell counts and cutoff values of the tills. Um and they have IIS tumor infiltrating lymphosytes the introumeral tails and perumemeral infiltrating lymphocytes pils. So the ones that went inside and the one that the ones that stayed outside. um and patients with high intratumerals and peritoumeral were defined um as those whose counts of both exceeded their cutoff. So they they decide on the cutoff and then they
00:12:36 - 00:14:01
decide okay which one is high which one is low based on the cutoff and patients with high showed better overall survival and recurrencefree survival. So these are the metrics that they check okay is something prognostic or not and then patients with low intratumeral and perittummeral teals um and after the um after the propensity score uh matching analysis. So they they they discovered they uh stated that this um these teal tails are an independent prognostic factor. So um the their conclusion is that they uh the
00:13:19 - 00:14:38
quantitative distribution of deals in esophagial cancer patients will help with the help of deep learning serves as an independent prognostic factor for these patients. Um and then they of course want to um do further work to decide if you can uh well to to check if you can um divide these lymphosytes into subgroup and subgroups and get like more granular spatial information uh to improve the predictive efficacy of tilts. This is something that's happening in breast cancer as well. H and basically like in the whole field of
00:13:58 - 00:15:02
immunoccology often you use different markers. We're going to have another abstract at the end like the classic of immunoconcology h immunoscore. H who has heard of immunoscore? Let me know in the comments. And I don't know if they're not coming through. Probably not. I'm going to see a bunch of comments on LinkedIn later, but if you're just joining, let me know where you're tuning in from and what time it is for you. And we're going to move to the next one. And I like the next one
00:14:29 - 00:16:03
because the the next one um is the the title is effect of liver biopsy size on MLDD fibrosis assessment by secondhand harmonic generation to photon excitation fluoresence microscopy. It's a mouthful but the important thing about this is that this is label free. It's a um like you image the tissue directly without processing and without imunofllororescent dyes right without anything and this uh is a group from UK Edinburgh London uh but also Bethesda and Richmond Virginia and whoever read on this topic knows that
00:15:17 - 00:16:33
fibrosis in metabolic dysfunction so this is MAS SLD is a metabolic dysfunction associated statotic liver disease. So it's the non I don't know if they don't call it non-alcoholic. Here is some like nash. Um basically this is caused or or this is associated with metabolic dysfunction not with alcoholic uh not with alcohol that causes fibrosis as well. So basically liver damage expresses itself if it's severe with fibrosis and this is a prognostic indicator and clinical trial efficacy
00:15:55 - 00:17:03
endpoint. So you're going to check if there is less fibrosis and then uh the drug. Okay, we have people from PA. This is so cool because I'm in PA and it's 6:18. Indeed, it's 6:18. Uh so cool to have you here, Jenny. Thank you so much for joining so early. You're amazing. Um, very amazing people who join at unusual hours and we not only have people joining very early but we have also people joining very late from different places. So uh and people joining from Europe for whom it is kind
00:16:29 - 00:17:59
of a lunchtime. So let me know if you're one of them and let's go back to the paper. So uh this um SHGTPF which is the second harmonic two photon excitation fluoresence microscopy um images unstained tissue sections. So it uses um fluoresence but it images unstained sections. So this is super cool because you don't have to stain right? So you know that staining and generating uh H& images takes like several hours, 8 hours, I don't know. Um I remember that we had to in my residency we had to switch um on all the
00:17:13 - 00:18:19
machines for staining and whenever somebody forgot it and like would come uh during the night because they remembered that they forgot to switch it on then the whole operations would be delayed for a few hours and we have guests from India joining the live stream. Thank you so much amazing to have you here. So anyway, we can skip that because we have we can image and stain sections. But let's get to the meat of this abstract. So when integrated with artificial intelligence models, they can generate continuous
00:17:46 - 00:19:24
fibrosis value, Q fibrosis and ordinal Q fibrosis stage. So um what they wanted to check here is the impact and so is the biopsy size and location uh is it important for accuracy um of this this thing right of the of the imaging on and of this diagnostic method. So they took one unstained section from uh 100 hepat hepatit or expplant masld cases um and 20 of each pathologist assigned non-alcoholic static clinical research network uh fibrosis stage uh okay so here we have fibrosis stage 04 so they
00:18:35 - 00:20:01
took 20 of each stage uh they used them to create virtual core um biopsies by cropping regions from within the whole parent section. So they have like a whole section and they want to make biopsies out of it. Um and they make them in different with different sizes but obviously the sizes that would correspond to the size of this biopsy in real life. Uh but they had varied length 5 to 30 millimeters um with a fixed width 0.9 mm um and fixed length 15 mm uh it's funny the regions varied in length with a fixed length anyway uh or
00:19:18 - 00:20:48
position uh within the whole parent section right and then the microscopy was used and Q fibrosis continuous value and stage of the virtual core biopsies were determined and what were the results my friends what did they say they say that this uh Q fibrosis so their AI based score correlated strongly with pathologist assigned stage so so these um these specimen already had an assigned stage and they checked if this Q fibrosis correlates with that and that was very high correlation uh um And then increasing the length and width
00:20:04 - 00:21:25
increase the correlation between Q fibrosis and a continuous value and the agreement with Q fibrosis stage stabilizing between 20 to 26 mm in length and 0.9 in width. So to a certain extent when the size was bigger it was the correlation was better but then after like you don't have to go bigger than that because then it doesn't make any difference anymore. Uh so conclusion longer over 20 mm and wider 9 mm meter uh 0.9 mm biopsies provide more accurate fibrosis assessment using the uh SHGTPF
00:20:44 - 00:21:47
uh biopsy position and orientation do not influence accuracy. So uh this is here kind of logical outcome because um you have the liver that is affected kind of homogeneously. This fibrosis is not going to be just like in one place and it's not going to be in the other place because it's going to it's like um a reflection of the liver function. It's not like tumor. You're going to have tumor in one place and the rest is going to be free. So if you take a biopsy here, you don't biopsy your tumor.
00:21:16 - 00:22:34
Whereas if you take your biopsy here, this is my biopsy needle, then you will have tumor and then like tumor heterogenity is a lot um more pronounced than the heterogenity of the fibrosis in the liver. It's kind of should be distributed homogeneously. So I would also not anticipate the biopsy uh position and orientation will influence but you know you have to check that because if it's your diagnostic modality you have to account for all these things. Um so what they suggest is that clinical trials should incorporate
00:21:54 - 00:23:22
suitable protocols to verify biopsy size to that optimize digital fibrosis assessment using our uh two photon excitation fluorescent microscopy. I like this one. I like because it's a um glass free I mean I don't know if I assume it's glass free like basically stain free and processing free way of imaging and last but not least let's talk about immunoscore give me a thumbs up or like is in the comments if you're well familiar with immunoscore how I am familiar with this immunos other than obviously reading
00:22:38 - 00:23:55
about it um in the literature. When I started my digital pathology career at um a digital pathology company, what we were trying to do is to develop an algorithm uh to quantify immune no they actually did it before I even joined. So um they worked with the original group original group that invented the immunoscore with Jerom Galon. Um and uh we're making a quantitation of this uh of the of the markers that come into the immunoscore. Uh and I see people joining here even 20 minutes in. Thank you so much. Let me know where
00:23:17 - 00:24:39
you're tuning in from. We have people from PA myself included from India. uh let me know where else you're tuning in from. Um so yeah, this is an image analysis based test the imunoscore. So let's dive into this one. Ah and in the meantime, of course, I'm going to give you the chance to check the store if you're interested in the earrings. Let me show you the earrings uh before we finish. So I have my multi-ucleated giant cells. Today uh we have three designs. Multi-ucleated giant
00:23:59 - 00:25:15
cells cartilage and um alium blue with um colon stained with alium blue that is like white and blue. So if you're interested in getting them for yourself or for somebody else, check this uh QR code that I have in the top left corner. At least looking from my perspective. Um let's do that. as I talk to you about the immunoscore redefining the light landscape of colorctal cancer and colurectal cancer control and care. Even more people joining. That's so cool. Thank you so much. Just looking at the chat. Oh my
00:24:39 - 00:26:06
goodness. We have people from Yemen, from Algeria. Amazing. Let me know what time it is in your places. Um cuz I kind of US I know and Europe more or less but the rest of the world I know Australia is like totally opposite wherever I am in the Philippines. So the immunoscore introduced by Jerom Galon in the early 2000s evaluates immune cell densities, keyword densities within tumors, um offering a more accurate prediction of clinical outcomes in colorctyl cancer compared to uh the tumor node metastasis. Sorry for the my drawings.
00:25:23 - 00:26:49
tumor the the normal like tumor node metastasis um system and this sorry ah nobody's telling me that it's not big let me know these things [Music] okay uh I got so excited showing you the earrings that I forgot to focus on here the goal of this live stream um so okay so we have the immunoscore that is a a better uh predictor than the normal system that pathologists were using and uh it enhances the prediction of recurrence disease-free survival and overall survival in early stage correl
00:26:06 - 00:27:22
cancer and uh identifying high-risisk patients and optimizing treatment decisions. So uh this immunoscore serves as a prognostic tool with high predictive value in guiding immunotherapy especially uh immune checkpoint inhibitors. So one of the famous one is going to be PDL1 and all these checkpoint inhibitors that are supposed to make our immune system uh actually fight the cancer um and uh like mobilize that's immuncology right our immune system um fighting the cancer and the drugs are trying to enhance that um
00:26:44 - 00:28:10
so high immunoscore score correlates with better response rates better response rates to uh immune checkpoint inhibitor therapy and then in addition to colurectal cancer it has been it has demonstrated utility across various cancers predicting treatment responses and survival outcomes. The original one is color colarctal cancer but also there are other cancers where they check this right and um it has demonstrated utility no sorry I said that utility uh across cancers but uh there are limitations so
00:27:27 - 00:28:44
the limitation is that it is an image analysis based thing right so if you don't have digital slides uh then you will be limited in the your ability to evaluate immunosore um and limitations with uh faces with traditional tumor infiltrating lymphosy assessments. Right? So you have the traditional one where you I you can stain them with IHC but also often without even staining. This one is uh specifically stained for CD3 and CD8. And then uh you calculate the ratios uh of these to each other and
00:28:05 - 00:29:31
in specific places right in the um in the invasive margin and similar to what we said with the um esophagial cancer. So the imunoscore application in guiding aduant chemotherapy is constrained by cost and limited data. So uh what they suggest is to advance this with combining it with other markers P21 and P16 that are markers for assessing cancer malignancy as well as integrating artificial intelligence and digital pathology. Uh but this score is based on well it's based on classical computer
00:28:49 - 00:30:21
vision way of assessing this lymphosytes but you can totally replicate it with image analog with AI powered image analysis. Um so yeah they say it's good but it could be better and it has limitations uh compared to normal um normal uh teals like tumor infiltrating lymphocy scoring h but they say it represents a significant advancement in oncology. So this was actually our last one for today. If you are new to this show, new to me, I'm Alexandra Zoraf. I'm a veterary pathologist passionate about digital
00:29:35 - 00:30:47
pathology and helping you understand it. And I also help you understand it with this book. And this is the QR code that I put on the screen right now if you're interested in learning more about digital pathology and you're just starting. Sorry, let me just I want to put myself on the screen. 27 live streams, guys, and I still like struggle what to click. Uh oh, look. I can put it here. How cute. I didn't know. So, anyway, the uh message I'm trying to give you here right now, if you're new,
00:30:11 - 00:31:18
get the book because you're going to have all the basics there and there's an AI chapter that's going to be uh updated in the new edition. and everybody who has the old edition. So, this is the physical book that you can get on Amazon and this is a paid thing, but the one from the QR code is uh free. It's PDF. So, you can totally get the PDF of this book, start your digital pathology journey, and whenever I have the next edition, you're already going to be on my list, and you're going to get the
00:30:44 - 00:31:43
book um the new version of the book, the PDF of the new version, and there's going to be a new version on Amazon if the earrings are something interested interesting for you. And also in this um store you can find other courses that we have a course about AI and about image analysis um in pathology. If that's something of interest to you, go ahead check the QR codes and I talk to you in the next episode. Thank you so much for joining me.