Computational Pathology:
The Next Generation of Companion Diagnostics

How will VENTANA TROP2 NMR RxDx Device* impact the field of pathology and what can labs do to prepare?

On April 28th, 2025, Roche announced that VENTANA TROP2 NMR RxDx, a novel computational pathology device, was granted FDA Breakthrough Device Designation, sparking a vital discussion around the state of current digital pathology adoption, the benefits of computational pathology, and what challenges and impact this new kind of companion diagnostic device will have on labs and
pathologists.

Dr. Landon Inge, and Mark Hellewell, experts from Roche Diagnostics Solutions and part of the team behind this upcoming device, discuss some of these challenges, explain the significance of the breakthrough device designation, and why AI shouldn’t be an afterthought in your digital pathology journey to ensure patients have access to these new, advanced testing options.

Landon J. Inge, PhD
Medical Affairs Lead
Global, Clinical Development and Medical Affairs

Roche Diagnostics Solutions

Mark Hellewell
International Product Manager
Personalized Heath Care Solutions - Oncology
Roche Diagnostics Solutions

How does the Breakthrough Device Designation (BDD) reflect the FDA and other global health authorities' current thinking about computational pathology-based companion diagnostics?

MH: Everything we're hearing from the FDA focuses on ensuring that this is an end-to-end device, to guarantee quality and reproducibility. And that's how we're approaching all our products now, which involves providing an end-to-end solution, from staining all the way through to image analysis. Obviously, there are many options on the market from a digital pathology perspective, but there's also a lot of variation between image types. And this can really impact the way a scoring algorithm could produce outputs. So, particularly in these early stages of computational pathology, it's important to show that we have this end-to-end device, which we can control every component to ensure that we have that quality.

LI: If you look at how any one of our companion diagnostic assays for IHC (not just Roche, but other manufacturers) are reviewed and approved by health authorities like the FDA and other regulatory bodies, they're always approving it as the end-to-end device. We provide data on everything from the primary antibody to the detection kit, the stainer, and the documentation associated with it. That entire device is reviewed by the FDA and other health authorities. That is because the same device is used in the clinical trial to collect data around patient enrollment, stratification, and patient selection. From the standpoint of the FDA and health authorities, all that the BDD and other computational pathology devices are doing is essentially extending that view to include the scanner, the software that the algorithm runs, and all of that becomes part of the device, which needs to be regulated as such. 

From a standardization process, we already know from our own use of the research-use-only** version of the BDD that standardization is extremely important. Once we start standardizing the pathologists' ability to score with these algorithms, we're now looking at the entire device being extremely important in regards to where we're seeing system bias and variability coming in.

What does it mean to be granted a Breakthrough Device Designation (BDD) by the FDA?

LI: The BDD basically enables us to have dedicated review and support from the FDA. It's not an approval. It's not even any sort of guarantee that we would have approval. It just allows for dedicated resources from the FDA to review our submission and provide us a faster turnaround time, much more interaction with the FDA about where the potential gaps are. But it also means that the FDA has recognized that it's a novel technology that needs that support.

MH: And I think it gives us that foundation to build from. We're not going into this blind with the BDD. It gives us that blueprint to drive forward with the end-to-end device, and that BDD really supports our vision for a complete computational pathology CDx system.

Why do you think pharma is interested in using computational pathology image analysis in clinical trials?

LI: I think a lot of this is due to published concerns around pathologist concordance for companion diagnostic IHC assays. For some companies, computational pathology represents a potential avenue for standardization and reduced variability of IHC assessment for clinical trials. And I think for pharma, if these improvements come in parallel with improved efficacy, that collectively provides substantial benefit for pharma. However, broad implementation of digital pathology still remains a hurdle for getting computational pathology used in routine clinical labs.

MH: It's also a signal about the way in which the market is starting to shift into looking at more in-depth insights as to what's going on within the tumor microenvironment—things that simply cannot be necessarily detected by eye, by the pathologist. And this is where we can hopefully increase the efficacy of therapy. We can identify new patient groups utilizing this type of technology. So I think it's certainly going to evolve, whether that's just in the space of IHC plus an algorithm, or maybe with multiple markers plus algorithms. I think it’s going to open the door for future opportunities within computational and spatial biology.

How is this computational pathology algorithm different from an aid-and-assist algorithm. Is it similar to what's on the market today?

LI: This device does not function independently of the pathologist, but it does function in conjunction with the pathologist. The pathologist has a well-defined role in working with the AI to generate the biomarker score. But ultimately, the AI and that computational pathology algorithm is what generates the biomarker score. So it’s not a standalone device, but it’s also not an independent device away from the pathologist. Both parts of the device need to work together to provide that biomarker score. It also provides something that the pathologist can't do visually. We're diving down into aspects of the AI system that reflect specific mechanisms, action, and biology of the marker that a pathologist could not assess visually, not just from looking at the slide, but also from the computational standpoint. We're not just calculating a score; we’re actually calculating scores on an individual basis of tumor cells, organizing that data, and determining the cutoff based upon that population data and where the quantiles and different cutoffs sit.

MH: Just to add to that, I think one of the aspects we need to focus on is that this doesn't replace the pathologist as the final sign-off. Where we differentiate between something like TROP2 NMR and an assist algorithm, assist algorithms are generally something that just counts cells. It's just a way of helping them be aware of looking at stain intensity or the number of tumor cells and staining molecules. It’s just that assist mechanism. Whereas with this one, it’s very much something the pathologist can't do by themselves, but they are integral to ensuring the correct area is scored and the correct tumor area is identified.

If the algorithm is generating the result, what is the pathologist's role in this device? Do they have input?

LI: They do have input. What we ask the pathologist to do is to function very much as a pathologist. The algorithm, at its basic level, has the ability to determine or decide on what it thinks is tumor or non-tumor. As part of that process, we actually ask the pathologists to review the algorithm's work in making that determination. The algorithm may not always be perfect in deciding what is tumor or nontumor or even what areas should not be analyzed. And so fundamentally, it's the job of the pathologist to check that the algorithm has selected the appropriate areas for scoring before that biomarker score is rendered. Again, it’s the pathologist and the algorithm working
in concert to generate the biomarker score.

Is a pathologist required to engage in this workflow? Can this be done by a tech or somebody without a pathologist's training?

LI: We've had a lot of interactions with thoracic pathologists in our study that we presented at World Lung this year in September. And, for these thoracic pathologists, their use of the algorithm is something that cannot be done by a technician or a non-pathologist. Some of the areas that are captured by the algorithm are the morphologies that really require the expertise of a pathologist to make the determination of what should be included or excluded. It's very much a morphologically driven determination of areas that need to be excluded or included.

MH: I think generally, we expect that the pathologist will be the one utilizing this algorithm. I think in certain countries, there are more senior staff who have a bit more role to play in terms of QC of stain, or some basic diagnosis. But generally, the pathologist will be the one guiding the use of this tool.

Regarding the data that was presented at World Congress of Lung, what was your main takeaway?

LI: One of the things that I would hope people understand with the data we presented at World Lung*** is that while pathologists still have a primary role with the technology, we can potentially make them perform better by utilizing these tools. We were able to show that, despite all the pathologists reviewing those cases independently, we could generate a concordance that was almost unbelievable because it was nearly perfect. If anyone has seen similar ring-type studies with PD-L1 or HER2 or any other biomarker, hitting almost 100% concordance is nearly unfathomable. I think we're really demonstrating what computational pathology can do for the pathologist. The flip side of that is now we need to focus on the other areas of the lab that we may have less control over but are equally important.

What does it mean to be a strongly supervised algorithm or device?

LI: This comes down to the level of involvement that occurs during both the training of the AI and the validation of the AI. From the standpoint of training the AI, this device is based upon every iteration of learning the algorithm has gone through. There is a substantial amount of work at the end of that training session, where pathologists go in and annotate, essentially correcting the algorithm's selection of tumor cells, nuclei, cytoplasm, and membranes to ensure that the AI is performing appropriately. Through those iterations of training and then essentially annotating or even checking the algorithm's work, that is what has ultimately evolved into the algorithm we have now. That fully supervised approach is much different from what you see with things like some of the generative AI that are currently on the market, where you just take an AI-based algorithm and you let it loose on all the available data that it can get its hands on. In that process, there's no sort of checking. It's very weakly supervised in the decisions that it makes. The outputs that are produced are all going to be based upon what it taught itself. No one is stepping in from a human standpoint to correct it or to improve its ability to disseminate true knowledge. So that's probably the biggest difference we have here. These fully supervised algorithms are really going to be central and critical for computational pathology, because we're not talking about a generative AI like ChatGPT where it's just generating information based on what it can scour off the internet. We're actually using this to determine patient suitability for drugs, so we have to make sure that it's performing appropriately.

What does a lab need to do to be ready for this device?

MH: A big factor in this is infrastructure. We recognize that, if we look at the IHC staining market, we almost know that we're going to find a Benchmark staining platform in most labs. A challenge we’ve got with the digital pathology market is that adoption is generally quite low. And it's quite fragmented. We see a variety of scanners on the market, as well as different types of image management systems. So it really goes back to ensuring that we have an end-to-end structure in place, which is the Benchmark staining platform, a DP 200 or DP 600 scanner, along with navify® Digital Pathology as the IMS. There’s also some additional guidance in terms of some of the hardware requirements, in terms of on-cloud or on-prem servers, to support the algorithm. But generally, it's about ensuring that they have the setup in place to begin the full range of testing.

LI: I think the other thing that labs need to start understanding and preparing for, outside of just infrastructure build, is from a laboratory standpoint, the quality of the IHC, the quality of the slides. All those factors that may seem routine for the lab staff are going to be
critically important to making sure that they're of high quality in order for these algorithms to work. There is already a fair amount of information out there about how variations in staining can throw off algorithms. So we, as a laboratory staff, are really going to need to be much more on top of how they're preparing those slides, monitoring the fixation and other pre-analytics associated with the tissue, in order to make sure that these algorithms can run successfully.

MH: I think it comes kind of full circle to something that we've been focusing on, which is around the education of the users, whether it's the pathologists or the lab staff. This is new for the market. So, gaining an understanding that we can't just throw anything onto the system and expect reproducible data to come off the slide. We really need to focus on the quality that goes in, and it will ensure that we get quality results coming out. I think a good comparison would be some of the more advanced testing that we offer on Benchmark platforms, such as HER2 and HER2 Dual ISH, where there is stricter guidance on pre-analytics to ensure the tissue quality is at the right level to expect optimal staining. So I think it's going to be the same type of education that we need to focus on here to ensure all labs are ready to operate with this new technology.

If a lab isn't able to invest in the digital pathology infrastructure, will we be providing any guidance on where they can go to send out testing?

LI: From the US perspective, we're anticipating this is going to be somewhat centralized at the beginning. Our plans, in collaboration with our partners, are to ensure that there are available sites for physicians and, more importantly, for patients to access this test. It is something that is definitely top of mind. I do think that there are other opportunities, based upon the infrastructure that we offer as Roche, to enable situations where, for example, some of the staining could occur at the site and even the scanning, but the analysis could occur in a different place. Or alternatively, the slide can be stained on site and sent to a different institution or place for scanning and the analysis. So we have built that into what we have currently within the BDD to allow for some of those decentralized, separate workflows, in order to help establish this as the first computational pathology device on the market.

What challenges will the lab face when implementing this computational pathology device?

MH: I know we talked about infrastructure, but one of the big factors is obviously going to be the access or reimbursement aspect as well. Right now, we don't feel the current codes that are available recognize the clinical value that Computational Pathology CDx devices offer to patients. That is certainly something that we’re working on and we’re expecting will come in the future. But in the early stages, there is certainly that concern about who actually pays for the algorithm and the device. Also thinking about the different stakeholders who will actually be using, or having impact or being impacted by the device. Making sure the lab understands the workflow about the workflow. Making sure that they are aware that this isn't going to impact them in terms of slowing them down. Another area we need to focus on is the oncologists, who may not be fully aware of this new type of technology. So, if we can educate both ends—from the staff all the way through to the oncologists—we can ensure that they have the knowledge and familiarity about what the capabilities of this tool are.

What is your role in developing this device?

MH: My role is currently global commercial lead for our computational pathology strategy. So, basically, overseeing all of our digital pathology programs which have a companion diagnostic algorithm associated with it, developing a commercial strategy, a go-to-market strategy, to ensure that we get as much adoption as possible and preparing the market for this new type of technology, and making sure that our customers and our affiliates are ready to support us.

LI: My primary role recently has been generating external evidence. From discussions with our pharma partners from the BDD there's a significant need for education, and particularly evidence for pathologists in the community to understand the technology. So right now, one of my major tasks has been generating external evidence with thoracic and other pathologists around the world. We just recently presented data at the World Congress of Lung Cancer with a research-use-only version** of the BDD. And this analytical study was really meant to support and explain to pathologists exactly how this technology works. We've been able to demonstrate a significant level of concordance utilizing the device, which really helps to build confidence within the community. We plan to do additional studies. As we get closer to hopefully having this device approved within the US, as well as other global markets, we, as a medical organization, will be providing as much training and education as we can so that everyone, pathologists, can feel comfortable with using the technology.

What does it mean to be a part of bringing this device to market? It's something novel, something new. We know that there's a potential for a huge impact on patients, and so what does it mean to you?

MH: For me, on top of everything you mentioned, it's always really motivating to work on novel technology and be at the forefront of new things that we're bringing to market. And with Roche, we always strive for that. But I think for me, it's exciting to look at it as something that could be really market-changing. We've seen digital pathology be around for a number of years, but without really getting any real traction, particularly in the clinical space. So I think this is really going to be a game-changer, not only for patients but also for the way in which labs operate and how they use digital tools as part of their day-to-day workflow.

LI: For me, it's something that I've been extremely interested in for a long time. I've always thought that computational pathology would be a game-changer for pathology and patient care in general. And I can really see this is just the first step into being what hopefully will be a larger deployment of digital AI into pathology, not just for IHC but for other markers. I'm extremely excited to be working on this as it’s novel and this is a first step, but I see this as a beginning to what hopefully will be a larger use.

*VENTANA TROP2 NMR RxDx is product in development and has not yet been commercialized in the US.

**VENTANA TROP2 (EPR20043) NSCLC RUO is a product in development and has not yet been commercialized in the US.

***OA09.03 Real-World Assessment of TROP2-NMR by Quantitative Continuous Scoring (QCS) in Non-Small Cell Lung Carcinoma (NSCLC), Lopez-Rios, F. et al. Journal of Thoracic Oncology, Volume 20, Issue 10, S29 - S30

Interested in learning more about our Digital Pathology Solution?
Fill out the form below to be contacted by a Roche representative