Event transcript

Read the transcript of this discussion below featuring Dr Marnix Jansen [MJ], Dr Maria O'Donovan [MOD] and Marcel Gehrung [MG]
 
[MJ] Good morning. Thank you. Welcome to this new series in the videos about Cytosponge and new diagnostic tool that will be introduced in the NHS for the early detection of Barrett’s Oesophagus, where we'll be discussing the findings from The BEST3 Trial as recently published in the Lancet. Here are Maria O'Donovan, consultant has two pathologists at Adennbrooke's Hospital in Cambridge, and Marcel Gehrung former PhD student with Rebecca Fitzgerald's developer of the Cytosponge and currently leading with cited on the commercial aspects of this product. We will be covering aspects of the histopathology as well as the sampling issues and the computer vision work that Marshall has been developing and how this impacts our patients. I'm Allan Johnson consultant, histopathologist with UCLH in London. I'm also a Cancer Research UK clinician scientists working on early detection or Barrett’s [Oesophagus]. Maria, do you want to kick us off?


[MOD] Great, thanks very much for that introduction [inaudible 00:01:16]. I'm going to talk to you a little bit about the pathology aspects of Cytosponge and also talk to you about the patient's Cytosponge story. I've worked for many years with Professor Rebecca Fitzgerald, where I've been focusing on developing the laboratory processes for the reading of the Cytosponge sample. Cytosponge you will be probably familiar with, is the pill on a string. The pill has got a capsule which encapsulates sponge. Patients swallows the pill, the sponge just released in the stomach. This is then withdrawn through the oesophagus and out through the mouth. It samples the entire oesophageal lining as its withdrawn. The cells that are retrieved from that then are suitable for biomarker assays and also for the development of risk stratification methodologies. That's what we've been focusing on over the past couple of years. There's the sample then it's a wonderful example because it samples the entire length of the oesophagus. We receive over a million cells in these samples. When you compare it to a biopsy specimen for example, here on the right-hand side of the screen you can see there are four biopsy samples taken during endoscopy. Those are two to three millimetres in size, whereas the scientist sponge sample is one and half to two centimetres in maximum dimension. They're far more cells as far more material to look out in that particular sample type. This is one patient's story, EC was 69 year-old female, she's retired research scientist and she had a 20-year history of reflux symptoms, and these were well-controlled with acid suppressants. She had previously been investigated. She had an endoscopy over 20 years ago. This was entirely normal and she had endoscopy on two occasions and both of these were normal. Her past medical history was not significant. She had had an appendectomy and that was it. She lived with her husband. She was a non-smoker and consumed very little alcohol weekly. She had no known energies and her medications then were the acid depressants or the antacids for her reflux. Because of her history of reflux and the fact that she was on these acid depressants, she was invited to take part in The BEST3 Trial, and she agreed to take part in it. She said, she was happy to take part because she wanted to support science. She also quite confident that her sample would be negative even that she had had two previous negative endoscopies in the past. She attended her GP practice, she swallowed the Cytosponge. The sponge was then placed in a fixative solution to preserve the cellular material for further studies. Here you can see the Cytosponge sample. There's the white flecks on the surface of the sponge or actually the cells that have been retrieved from her oesophagus. The sample was then sent to the laboratory for processing. These cells removed from the sponge and then they're concentrated, you can see them in this yellow area at the lower end of this conical shaped tube. They're, process and formalin fixed, then embedded in this paraffin wax. Then very thin sections are cut and stained with a Hematoxylin and Eosin stain and ready for then examination under the microscope. The oesophagus we would expect to see when the biopsy you see the normal squamous lining, which is just to these layers of cells which are entirely normal. On the right here you can see the lining form a patient who Barrett’s Oesophagus, you loose that squamous lining and it's replaced by these columnar or rectangular shaped cells that form the ribbons of cells. They contain these spaces which are filled with mucin releasing goblet cells. These are indicative or intestinal metaplasia. These are characteristic finding in Barrett’s Oesophagus and bears an endoscopic image during relation coronary. The white area is the [inaudible 00:05:47] and the red areas is looking down into the oesophagus, down to the normal of the oesophagus. The red area is the area of Barrett’s Oesophagus. This was bursitis bunch sample, it was a very cellular sample, lots of cellular material present on all of these polygonal shaped structures. The pink structures here are the squamous cells. These are entirely normal. There has been some normal oesophageal sampling. But in addition, she's got this ribbon of cells, this strip of cells which are columnar shapes. They're rectangular shaped individual cells. They have these rounded goblet cell structures which are areas of intestinal metaplasia. This conference that she actually has intestinal metaplasia and would be in keeping with Barrett’s Oesophagus. Then we also perform an immunohistochemical stain check on factor three or TFF-3. This highlights these goblet cells and this brown colour so that we know what the this is a focus of intestinal metaplasia and keeping with Barrett’s [Oesophagus]. She also had evidence of ulceration. She had this acute inflammatory exudate with Fibrin and inflammatory cells. In keeping with an officer.
She was then invited for an endoscopy based on the positive Cytosponge findings on endoscopy. This is, again, looking down through the lumen of the oesophagus and this is the lower end which shows the stomach. When we're really looking at the inside lining of the oesophagus, we can see there's this red colour. Slightly irregular, bumpy surface. It should be nice and smooth, it's slightly irregular and bumpy which is not normal, and also has this yellow area of ulceration on the six o'clock location. On autofluorescence imaging, the normal autofluorescence imaging colour is this green colour and this is last in her endoscopy. You can see that it's demonstrating a pink colour and the lower end here. This is this pink colour, and this is abnormal. This is a characteristic finding in neoplasia or early cancer of the oesophagus, when you lose the normal colour. She had a biopsy taken in this demonstration parts. There was high grade dysplasia, which is premalignant conditions and also an intramucosal adenocarcinomas so there was an early cancer of this site. Her case then was discussed at a multi-disciplinary two meeting and it was decided that [inaudible 00:08:20] was localized and superficial within the oesophagus, then she would be suitable for endoscopic treatment only and would not require an esophagectomy. She had an endoscopic ultrasound and she also had a CT scan. The ultrasound showed some thickening of the mucosa at the lower end of the oesophagus. There were no abnormal lymph node stem [inaudible 00:08:42] PE scan to evidence of region on our distant metastatic disease. It was decided that she should have an endoscopic resection of the inside of the oesophagus in the abnormal area. During the endoscopic resection, this is the [inaudible 00:09:00] beforehand where you can see this nodules, they are bumps in the lining of the oesophagus and after the procedure then this is what was left where you were left with this red area where the lining has moved and we are now looking down onto to a middle layer of the oesophagus and muscle layer of the oesophagus. The oesophagus is then sent to the laboratory, and this what we saw under the microscope. The normal oesophagus is just in the corner here for comparison, the normal lining but instead of this patient, she had an abnormal columnar epithelium of these abnormal glandular structures, and then it will become very irregular, they're crowded, they're at all different. They're spaced irregularity. When you look at them closely, you can see there are abnormalities and there are some variations in the shape and size of these glands and they're extending deep. This is superficial and it is extending deeper within the tissue. We're still just extending beyond the mucosa when we examined the whole tissue resection, we were able to identify the depth of invasion and it was superficial, it was just involving the superficial part of the submucosa. The deep margins were clear and there was no evidence of vascular space involvement which would indicate a worse prognosis. It was present. She had an endoscopy five months later, and there were some good healing. The white areas here are the squamous, we have epithelialization or squamous healing. There were still some red areas of pharynx and there was exactly irregular area with a little bit of abnormality on the right-hand side of this area and our narrow-band imaging, this showed again slight modularity and also some variability in the vascular classroom at this site. This was biopsied, it again showed resistant high-grade dysplasia and also a focus at the suspicious for intra-mucosa adenocarcinoma. It was then decided that she should have a further endoscopic resection of this area. She had that removed, and again it confirmed a superficial early cancer and emotionally differentiation in the mucosa adenocarcinoma This time it was more superficial than the last one. Her margins were again clear and there was no evidence of vascular space invasion. She's been followed up two years since her secondary infection and she's been completely disease free with no evidence of residual Barrett’s [Oesophagus], no evidence of pre malignancy or any cancer. In this case, the textbook as well tolerated. It is certainly a cost-effective screening device for pre-malignant and early cancerous conditions and this lady had successful screening with Cytosponge. It led to a diagnosis and also to minimally invasive treatment of her early oesophageal adenocarcinoma or early oesophageal cancer and ultimately to a cure. Thank you.

[MJ] Thank you Maria. That's a lovely example of early cancer detection, whereby this screening tool effectively has also been used, which is really the goal for the whole method, of course, towards early cancer detection. Marcel, Maria and I have seen many, many biopsies of these patients is of course a completely new approach to screening the premalignant oesophagus. I was hoping you could give us an introduction to your work on the computer vision aspects of this.
 
[MG] Great, thank you very much, Marnix for the introduction and Maria, for the first part of the talk, and I would pick up some of the topics which Maria has touched on as well. In the next few minutes, I would want to give you a brief overview on the use of artificial intelligence and how we can support this pathology process, which Maria has demonstrated for Cytosponge, Chapter 3 for the early detection of Barrett’s [Oesophagus] oesophagus and eventually adenocarcinoma. To quickly recap actually on something which Maria has presented as well. One of the main challenges we have here and also one of the opportunities we have here is that this is a fairly new sample type. Over the last many, many years, the team has built a very, very thorough understanding of how pathologists are approaching this particular sample type as well, and what our or my project of work was basically on, is trying to understand this process in depth and building a computational framework which mimics these heuristics how a pathologist approaches this type of sample. To give a brief recap on what the data is, we have to work with is. This is what Maria has shown as well but this is basically what happens when Cytosponge is withdrawn from a patient sent to a laboratory. We process the Cytosponge and we get two slides with two different stains for one patient sample. One of them stained with haematoxylin and eosin, and the other one is stained with Trefoil factor 3. Why those two different stains are important and particularly important for how we deal with the state of from a computational perspective, we will have a look at in a minute but before we do that, I actually want to give you a brief insight into why we think that we need to develop assistive tools for pathologists to read Cytosponge-TFF3 samples. To do that actually, a couple of numbers here has a demonstration. Over the next couple of years until 2025, we hope to roll out this technology into the wider NHS and some early estimates here concluded that we would aim for around 300,000 estimated procedures per year. If we do some simple math here and assume that it takes us about 8-10 minutes of the pathology where you are one of the samples we get to a very, very large number of 50,000 hours of pathology reported required per year, which if we put this in line with the number, for example, is pathologists available in the UK for that type of work, we really hitting the boundary there, so this is something which needs to be urgently addressed and hence it's something we started to work on a couple of years ago. The specific toolkit or a set of tools which we use to do that is called computational pathology, which basically means using computer vision tools on history or psychopathology images and trying to extract meaningful information from these images and then prepare or curate the information which you get out of this computer vision process and present it to pathologist or interact with the pathologist in a way that it helps them to, first of all, increase the screening speed of those pathology samples, but also doing that without interfering with the diagnostic accuracy. How you can do this is in many different ways, but to group it into the two most important abstract concepts, one of them would be full automation and if we take this in the context of Cytosponge, what I mean here is if you look at the left here, we, for example, have a deep learning classifier which has been trained on individual regions or images of cells, which is able to look at the H and E stain and the TFF3 stain and is able to get us in probability for what type of cell is present in a certain area of the image. Why making it a distinct differentiation between the H and E stain and the TFF3 stain here, that's because I'm after working very closely with Maria Donovan and her colleagues and the pathologists who are very familiar with reporting the sample type from a computational perspective, we basically divide the analysis of Cytosponge into quality control, which is mostly referring to looking at the haematoxylin and eosin section where we are trying to find gastric epithelium to check whether this sponge has travelled all the way into the stomach and the TFF3 bio-marker, we're using for the diagnosis to check for the presence of intestinal metaplasia. To go back to this automation approaches which I show here, what you can do is you extract, for example, these information from the images. You get some estimate, for instance, certain cell type present we need for quality control and diagnosis, then you do a fully automated classification and you generate a report. Big problem about this is that there's actually no pathologist involved in that but one of the things which you will see now it's a semi-automatic approach which we actually want to leverage by keeping the human in the loop here is that machines are a lot of computer vision are quite good at very easy images or very easy examples and make a call on them but actually the harder these images get, the more we would benefit from keeping the human in the loop and making a judgment on these cases. What we can do here, for example, is we can actually extract additional statistics from these deep learning processes around how confident the algorithms are in making a certain decision and then we can take the patient samples where the algorithms have lower confidence, making an automated call, giving that to a pathologist, to an experienced pathologists for reviewing them while all of the unequivocal samples would then be reviewed in an automated way, and thereby reducing the workload and giving only the cases where we really meet the human observer for the pathologist and thereby reducing the workload.
 As I said, we developed this framework and put it in the context of Cytosponge as well, and developed it, as I said, with these two different processes in mind, around quality control of finding gastric columnar epithelium on H and E sections and finding positive goblet cells. It stains brown round or white cells which Maria has shown for the diagnostic intestinal metaplasia and then stratifying those patients’ samples into individual confidence glasses. Then pushing the ones where we have lowest confidence, to pathologist for manual review. How this looks like if we look at some data here you can see it's actually for the quality control task at the top you can see whether a pathologist has considered the Cytosponge sample as passing quality control or failing quality control. At the bottom you see the number of areas in the image which might be indicative for this patient passing quality control extracted by an algorithm. What you can see here, the further this number of detected relevant regions declines to what's the right of the image, the lower the confidence of the algorithm gets. But also at some point where you see the low confidence bracket, it doesn't agree with the pathologists in all cases anymore. All of these patients would benefit from having had it reviewed by a pathologists instead of going through automated view. But you see that the algorithm actually performs very well for these high confidence and no confidence regions on the left and right side of the image. We can now do the same for diagnosis except that here we can compare ourselves against two different layers of ground through. The first one is, what did the pathologists say on the basis of looking at the scientist's TFF3 sample. Also then, because this is some data where we had endoscopy data for the patients as well, what was the endoscopy finding for this individual patient? Again, you can see that for the high confidence regions on the left and right, we actually get very good agreement between the automated assessment, which you can see at the very bottom and what the pathologists says and for the low confidence region in the middle, we can clearly see who benefit from having an experienced pathologist looking at these images. Take one extra step now, how this is basically implemented and how we validate this as well, is by running this on two very large validation datasets. We basically ran all of the sample through an automated pathway, except that we only took the automated analysis or automated result of patient samples and use MSD as the final result if they fell into high confidence categories. We took and substituted the pathologist results into all of the ones which fell into the low confidence categories. To summarize that, we call this a triage-driven approach and one of the main goals here for us was to reduce the workload and we could actually show in two validations that we were able to reduce that workload by 57 to 66 percent. We've validated this approach in over 4600 slides from 2300 patients and something which was very important to ourselves as well, and it's something I referred to earlier about how we could use computational pathology for these types of samples. We also were able to ensure that we had stable diagnostic test performance, which means we have no decline in sensitivity or specificity and only very minor decline or decrease or increase at the same applied for using this approach on the best three study where we obviously look at positive predictive and negative predictive value. Exactly that stable diagnostic test performance was insured because we were forwarding a few of the cases for manual review by pathologists. Thank you very much.


[MJ] Thank you, Marcel. I think that's a lovely overview of how you've pushed forward on the computational pathology there, analysing the samples and also highlighting how this completely new approach whereby you've used both the comparison with the pathologists diagnosis on the sample as well as the eventual endoscopy diagnosis to really get a better understanding of the performance of the scientists bunch readout. Maria, if I could just return to you because I think that's an important point, as I said before, you and I have seen many biopsy trays whereby the standard methods still of screening patients that are undergoing endoscopy for Barrett’s Oesophagus. Biopsies are taken every two centimetres along the length of the Barrett segments. We get four biopsies at each height and this is of course a completely different approach. I was hoping you could say something with regards to the difference in sampling and also possibly the heterogeneity that you might see comparing biopsies to this cytology approach?
 
[MOD] Yes. Yes, Marnix. That's one of the big advantages of this particular technique. It's that you don't need to look at multiple biopsies all your oesophagus is sampled in the one specimen. Until the advantage there is that if there's an area that hasn't been identified by the endoscopist as abnormal then the Cytosponge will pick that area up. It increases the chances of detecting any abnormalities within the oesophagus because you're not entirely reliant on the endoscopists identifying an abnormal area. Also you have one sample to look at, so it reduces the amount of time that it takes the pathologist to look at those samples. Because if somebody's got a very long segment of Barrett’s [Oesophagus], we're talking about a nine-centimetre segment of Barrett’s [Oesophagus] or ten, that's a lot of biopsy samples if you're going to have four biopsies for every two centimetres. Whereas one Cytosponge sample, will sample the entire area. We have to of course be mindful that the Cytosponge not alone, just sample the oesophagus, it goes into the stomach. You can get the upper end of the stomach samples, but also the mouth area, the oral pharynx area is also sampled. It's important that pathologists are then trained to interpret and to read these Cytosponge samples because you have to make sure that you're confident in the cell types that you're looking at and that you can confidently say that what you're looking at is from the oesophagus or stomach are rather than from the oral pharynx, the area of mouth. That does come into play.


[MJ] Yes. I think for the histopathologists who are still pathologists watching this, an important question will also be the technique is really rather simple. I've also seen in patients whereby they swallow the pill then wait a couple of minutes and then the sponge is retracted. All of this can be done on an easy outpatient basis. Can these slides also be read in any histopathology lab in the UK?


[MOD] No, they can't and the reason I say that is although the processing methodology is relatively straightforward, in order to make sure that we have consistency of the sample type, consistency of its staining that we can audit the results, it's important then that it's done in a central laboratory. Also the training of the pathologists, as I mentioned already, is really key to this because you don't want to have over or under interpretation of the samples. It's really important that there is a unified pathologist group that's looking at these samples. Also helps with quality assurance. We have an external quality assurance scheme set up for this so just to make sure that everybody is reporting in the same way and the reports are standardized. That's very helpful for the clinicians that we issue standardized reports. It's better to have centralized processing and reporting for the sample type.


[MJ] Yeah. Because obviously that's one of the main caveats currently to Barrett’s [Oesophagus] diagnosis and also to dysplasia diagnosis is the subjectivity of the diagnostic. Marcel, I was hoping you could say a word to this with regards the future developments in this area for computational pathology in the pathology lab or how this impacts the daily work of pathologists.


[MG] Yeah, I think it's actually quite an interesting question, especially in the context of Cytosponge as Maria just mentioned as well we have a very, very well constrained process which enables us fairly easily to work with pathologists to target individual problems they're countering or bottlenecks for time, for example, which they encounter with this particular sample type. But more generally, I think there are three different areas in which computation pathology will impact those pathology workflow. The first ones are basically solving workflow problems or solving approaching these things for workflow perspective. For example, when Maria our colleague or you as well look at pathology images, making sure that your focus or your vision is directed to the regions that actually matter in the images. For example, one of the ways how these pathology tools or computation pathology is used these days. Then there's another completely new category which is something we are exploring on Cytosponge as well, which has more of the prognostic and predictive characteristics. Can you see more than what the pathologist, for example, sees in these images as well. Is there a relationship of the underlying morphology for which something we are not really aware of, that there is actually a relationship at hand there. How can we build those tools to disentangle that and make that information available to the conditions for the guide that patient care pathway. But yeah, so those are a couple of the different perspectives of where our technologies can make an impact in the coming years as well and obviously particularly in the context of Cytosponge as well.
 
[MJ] Maria, back to you. The method is designed as a screening method for Barrett’s Oesophagus. Have you been thinking about using this towards other pathologies as well? Are you thinking of patients using it for esophagitis, of course, a condition that's becoming more common as well?


[MOD] Absolutely, yes, we're indeed. We've been seeing examples of eosinophilic esophagitis in Cytosponge samples during our trials, and as I mentioned with the patient who I discussed in my talk, she had an area of ulcerations so we're identifying inflammation. Sometimes you can have viral infections, herpes, etc. Occasionally can be seen in the oesophagus and we have identified that as squamous abnormality so abnormalities of the normal squamous lining where you can get squamous cancers and squamous atypia. We're certainly seeing those as well. We've identified Helicobacter pylori, which you can see in the stomach. We've seen that in some of the columnar epithelium that's come from the stomach. There are a range of abnormalities that pathologists see in the oesophagus that we do detect in Cytosponge.

[MJ] Yeah, that's really nice. Marcel, are there any ways in which we could think of applying current biomarkers or biomarkers that are still in development to this technique?
 
[MG] Yeah, I think there are different aspects, actually. One of them related to the point which Maria also just made. There's a lot of information in the H&E stain which is hidden in the morphology or in ever so subtle changes and even if they're very under-represented, and with respect to the thing you are actually looking for, that's where computation pathology tools can actually help because they are objective when it comes to looking at the entire image, and they can highlight those regions, and they might actually show and highlight that region which a pathologist can miss. You can think as that computational ecology toolkit as almost like a bit of a biomarker, which can help focusing a pathologist on a specific region. But there's also other things which we have been working on in the last few years as well and maybe that's something Maria you want to comment on as well. Working for example, with p53 in the context of very specific patient groups to further guide stratification of these patients, especially if they come from a different target population than the one we targeted in Best3. Maybe Maria actually you want to comment on that, how we can use P53 to support the certification of these patients?

[MOD] Yes, absolutely. We're using p53 routinely now with patients who we know have Barrett’s Oesophagus and we're using it to help guide management in terms of their risk. It is well known that if you have aberrant expression of p53 then you are at increased risk of progressing to neoplasia of the oesophagus. So we're now looking for that aberrant expression in Cytosponge samples in patients where we know they have Barrett’s Oesophagus and they're being screened regularly for their Barrett’s Oesophagus to see if they're showing any signs of progression to dysplasia or cancer. This is a very useful biomarker. It's also useful in the context of squamous dysplasia because the squamous dysplastic cases tend to show aberrant p53 expression as well. So it's useful in that context too. Although we see less squamous abnormalities here in the UK, but we have seen certainly increased numbers of squamous abnormalities in other countries because we have few cytaspongia and other high-risk squamous cancer areas for example, in Iran, in China and Africa.

 [MJ] Yeah. I think finally really, so this is really a beautiful conjunction of this new screening methods to identify or kind of methods to identify more Barrett’s [Oesophagus] patients early. Marcel, you beautifully illustrated your cutting-edge computer vision work that's applied to this method. For the interested patients or pathologists out there who are currently looking at Barrett’s [Oesophagus] biopsy, so where this is currently stand, what's next for the Cytosponge, and what phase of development is this really?


[MG]  One of the things we have seen over the last year, and this was very much fuelled by the coronavirus pandemic as well, is that the restriction in the availability or the limitation in the availability of health care services has put a lot of pressure on the endoscopy services as well, so the uptake and interests in Cytosponge has grown rapidly over the last year as well with large projects being rolled out. In Scotland right now, with NHS Scotland and also with NHS England and we're focusing on right now to actually support those limitations of endoscopy service availability by deploying the technology in secondary care where it's most useful at the stage. It basically helps reducing the waiting lists in hospitals so patients can receive access to care sooner because [inaudible 00:35:03] it's a procedure, it's something which comes with less patient discomfort, and that's something which can be done in an office setting. It does not require as many people in the same room, and for example, exposure to aerosol than an endoscopy procedure needs.
 
[MJ] Maria, would you like to add to that still?

 [MOD] Yes. We're currently, as Marcel has said we are currently receiving samples from NHS patients, so this has now moved to the clinical setting, and that's very exciting for us. Certainly, the numbers are increasing week on week.

[MJ] Yes. That's also what's happening in my Trust. Marcel, as you've outlined this with the pressures on the endoscopy department with COVID clinging in between endoscopies, this has really fuelled the need and we've now also started screening patients in this way. Well, thank you both. I think this was very interesting. If you're watching this, I would urge you to look at the other videos as part of this series as well and thank you very much again.