How I AI

How a Doctor and Scientist Uses AI in Medical Imaging and Diagnostics

Brooke Gramer Season 1 Episode 19

This week on How I AI, I am joined by Dr. Akash Parvatikar, a computational pathologist and medical imaging scientist who works at the intersection of AI and healthcare.

Akash earned his Ph.D. through the joint Carnegie Mellon–University of Pittsburgh School of Medicine in Computational Biology and has been developing advanced AI workflows that support doctors in analyzing medical images and improving diagnostic accuracy.

We explore how large language models (LLMs), Python, and agentic workflows are being used in modern medical imaging labs, the move toward digital pathology, and the new possibilities that come with remote collaboration and telepathology. This episode offers a behind-the-scenes look at how scientists are approaching AI adoption in ways that are practical, regulated, and patient‑focused.


🔥 Topics We Cover

  • How computational pathology is shaping the future of medical imaging
  • The shift from physical slides to digital workflows in pathology
  • How AI tools assist specialists by streamlining repetitive image analysis
  • Agentic workflows and why they are not just for large language models
  • The role of telepathology in connecting specialists globally

🛠️ AI Tools and Workflow Akash Mentions

  • Python for building AI scripts
  • NumPy for numerical computing
  • Scikit-learn for machine learning tasks
  • LLMs and agentic workflows for diagnostics
  • Digital pathology platforms for slide digitization and quality review

📲 Connect with Dr. Akash Parvatikar


🎧 If you enjoyed this episode, please leave a review and share it with a friend working in healthcare, AI, or life sciences. How I AI is here to spotlight the quiet builders and brilliant minds using artificial intelligence to elevate real human impact.

Want to be featured?
Have a compelling AI use case or experimental workflow? Apply to be a guest at howiaipodcast.com.

Interested in The Collective? Skool gives you immediate access to challenges, academy webinars, and community. 

If you’re ready for a more immersive experience, like done for you solutions, high tier memberships, or in-person mastermind events, reach out and I’ll share my exclusive Collective AI offers

More About Brooke:

Instagram: thebrookegram

Website: brookex.com

LinkedIn: Brooke Gramer

More About the Podcast:

Instagram: howiai.podcast

Website: howiaipodcast.com

Learn more about Brooke's exclusive Collective AI offers.

"How I AI" is a concept and podcast series created and produced by Brooke Gramer of EmpowerFlow Strategies LLC. All rights reserved.

Brooke:

Welcome to How I AI the podcast featuring real people, real stories, and real AI in action. I'm Brooke Gramer your host and guide on this journey into the real world impact of artificial intelligence. For over 15 years, I've worked in creative marketing, events, and business strategy wearing all the hats. I know the struggle of trying to scale and manage all things without burning out, but here's the game changer, AI. This isn't just a podcast, How I AI is a community. A space where curious minds like you can come together, share ideas, and I'll also be bringing you exclusive discounts, free trials and insider resources so you can test drive the latest tools and tech yourself. Because AI isn't just a trend, it's a shift. The sooner we embrace it, the more freedom, creativity, and opportunities we'll unlock.

How I AI is brought to you in partnership with The Collective designed to accelerate your learning and AI adoption. I joined the collective and it's completely catapulted my learning, expanded my network, and show me what's possible with ai. Whether you're just starting out or looking to refine your AI strategy, The Collective gives you the resources to grow.

Brooke:

Stay tuned to learn more at the end of this episode, or check the show notes for my exclusive invite link.. Today on How I AI, we are getting an insider look at artificial intelligence from someone who sits right at the crossroads of medicine, science, and technology. Dr. Akash and his work at HistoWiz in New York is changing the way we diagnose cancer. Akash has a front row seat to how AI is being adopted across industries and has a superpower with helping teams figure out which technologies will actually move the needle for them. If you are curious about the real advantages of how scientists are adopting AI and what that means for making smarter, more intentional tech decisions, this episode is for you. Welcome to another episode of How I AI I'm your host, Brooke Gramer. Today I have a very exciting guest. His name is Dr. Akash Parvatikar. He's a scientist who works with computers and medicine to help doctors understand diseases like cancer. Dr. Akash. I'm so happy to have you.

Akash Parvatikar:

Thank you so much, Brooke, for inviting me here and I'm sure we'll be having a lot of interesting conversation. Right now I'm working at the intersection of AI cancer diagnosis to help clinicians and researchers to accurately diagnose and also help them diagnose more quicker with the help of AI. Once again, thank you so much for having me here,

Brooke:

yes. I was so excited when you reached out expressing interest to be on the podcast because something I actually didn't share with you in our intro call yet I've had three very close family members be affected by cancer. So when you shared a bit about your background, I was so interested to dive deeper of this. Intersection of AI and medicine and AI for good because there are a lot of negative connotations that are happening right now with people adapting AI, and I'm so happy to shine more positive light onto this topic. And so let's kick it off here,

Akash Parvatikar:

Sure.

Brooke:

I'd love for you to just expand on your background and how you got into medicine and where you're at now with, with AI?

Akash Parvatikar:

Oh yeah absolutely. So I did my bachelor's in electrical engineering then got my master's in information science. And during my master's I was doing a research project. In the com bio department at University of Pittsburgh, where I was the first time exposed to the biological dataset for drug discovery. And then after doing the research work at the University of Pittsburgh, I joined the PhD program at the Joint Carnegie Mellon and University of Pittsburgh Medical School. In computational biology wherein the problem at hand was to understand why is it difficult to diagnose preinvasive breast cancer, and why is there so much of disagreement among the doctors in the disease diagnosis? So for me, I, A wanted to understand, at the primary level why is it so difficult? And B, being a computer scientist, how I can build solutions for the domain expert who are pathologists to help and solve that problem. So for me, it was more problem driven, which is the cancer diagnosis and then, building state of the art solutions to help both understand the problem and then build solutions on top of that.

Brooke:

Wow. A lot to unpack there. I have a standard set of questions I ask everybody, but it's different every time based

Akash Parvatikar:

Yeah.

Brooke:

level of integration. And with you, you're on such a high level. My first question is, and maybe you can break it down to an easily digestible way so we can understand it. What kind of tools and machine learning do you use with your role? What is your quote unquote tech stack look What kind of technology are you using to do your job?

Akash Parvatikar:

Right. So while I was doing my PhD, and this is a little over four years back, so I had to build my own frameworks from scratch. So all that I used was the Python programming language and few packages NumPy and Psychic Learn and all that to load the data. But I built my own framework, my own objective function and all that. And that gave me a very inep understanding of what it takes to build AI. For medicine and also build explainable and trustworthy AI, which is so important because, yes you could use the solutions out there, which is already available. But when it comes to regulatory approvals and having trust in the AI system the clinicians want to know, if you're using any open source software, what it was trained on, where did the data come from, who did the annotations and all that. So far, my phD work. It was from scratch. Did not use any open source software. But right now where I'm currently working at HistoWiz we do use some open source software for disease detection and all that. And it's we, we use the transformer networks, we use unit for segmentation. And we have been sort of venturing into these LLM models as well. How can we build agent AI to solve very specific tasks. So it is quite diverse within the company that I work for. But I think the, tech stack it's problem driven. It depends on what you're trying to solve.

Brooke:

Maybe you can dive deeper on what is changing right now in the medical space when it comes to implementing AI and specifically in your field with cancer diagnosis.

Akash Parvatikar:

Right. So what's changing? There's a big change happening actually. So up until recently, and even now, as you might know, that the pathologist look at the physical glass lights under a microscope. So these are non digitized. These are physical raw slides that they use microscope to look at the tissue and do the diagnosis. But then what changed in 2017 is that US FDA approved using digital scanners for primary diagnosis. So what that means is now you can take these large resolution images. Of your physical glass slide and a clinician could look at the digital image at a high resolution. So that opened up a Pandora of like.. Good opportunities to share the data quite easily implement AI and do your further analysis. So so that's what's changing, which is going from manual anatomic pathology to digital anatomic pathology. So that's the switch that's happening within my field. So say the slides are already digitized. There have been a lot of solutions out there to run AI for diagnosis, but I think that's where there are some trust issues between the AI developers and the pathologist. So if the AI is directly diagnosing a particular case, because a cancer diagnosis is so complex, that is, there are so many things happening within your biopsy tissue and no AI system is is able to accurately detect cancer for you. But what's happening right now is AI is doing a really good job in those tasks, which takes a lot of time. For the doctors to detect, for example, if they're looking at, some very specific regions within the entire image, now AI can do a very quick job in doing that and also quantifying things for you. So right now all the AI tools in my field or in what's coming up is these assistive tools than decision making tools. That's coming into play. For cancer diagnosis, again, I'm more than happy to expand upon what I said, but this is where the field is going right now.

Brooke:

So you are expressing that we're all moving to digital, which is saving a lot of time, and you're able to find the issues in the tissues more accurately and So my next question, because you chatted a little bit about the trust factor when it comes into this use of AI in medicine. Can you maybe expand on that?

Akash Parvatikar:

Sure. I think not just in medicine Brooke. And I the way you said the issue in the tissue, using ai, I think that that should be, I don't know, a tagline for one of my posters or something. But so talking about the trust factor, I think not just in the field of medicine, any AI adoption. There is this hard approach. Where AI takes the decision for you. And there is a soft approach wherein, you could still use the AI in your routine tasks, but you are the final decision maker. And, you could use it as a tool to better help in your own workflow, then create a whole new workflow where it's completely done by AI. So in the field of pathology, what's happening is that say that AI gives you a 90% accuracy of detecting, say, breast cancer. But then for FDA to approve that AI system, we should absolutely be sure of those 10% cases where the AI does not do a good job because if we do not know that, then we have no idea when a patient will be wrongly diagnosed. So it's okay if the AI does a good job, say 80% of the time, but we should be very sure what are the failure modes of this AI system when the AI does not do a good job? It could be the race of a woman undergoing a breast biopsy. It could be the region where they came from. So the AI was trained on the population of US and Europe and say the patient belongs to West Asia. Will the AI do a good job? So I think in facilitating a regulatory approval, all of this has to be accounted for. Because it's good to know, okay, fine. This is where the AI does a good job, so we might just implement over there, but we cannot use the validation report for a certain use case and expand it worldwide and make it a general solution. So, especially in the field of medicine, there are very, specific AI use cases that's coming into product uh, mode, and even the regulatory approvals are given for that specific case. Up so much that, if some AI is trained on, say, NYU data and they get an FDA approval, so they could treat only NYU patient with that AI system and they cannot go to say Columbia or, university of Miami Hospital and all that. So I think it's highly specific.

Brooke:

And you touched on a use case just now. Could you maybe share a couple use cases, maybe one or two of success you've had integrating AI into your research and into your work.

Akash Parvatikar:

Sure. I think, uh, the first large success that we observed is integrating an AI to detect quality issues.

Brooke:

Hmm mm.

Akash Parvatikar:

In the data so we, at HistoWiz, we produce tissue data from multiple organs and species. And for us it takes a lot of time to detect quality issues within the image. So say a particular image could be bloody, or it might have some artifacts on there. Or it could have some folds and breaks and all that. Now, having a person trying to look at thousands of images, identifying these, I don't know, pen marks, folds, blurs, and all that, takes a lot of time. Because we are dealing with a very large sized image. So we have developed and integrated a novel AI solution. To detect a very simple thing, which is quality issues in this imaging data because the development of AI cancer detection and all that comes next because if you don't have a good quality data. It's garbage in, garbage out. So you cannot trust what the AI is doing, further on. So now that's one big success that we, we have got. And we are also having very interesting discussions with not just other AI companies who want to use our AI tool for Quality control, but also from College of American pathologist, which is CAP. And even at the federal level they're interested in using our quality AI tool because we have been using it every single day within our company. And we have a large clientele and, to produce good quality data. So that's a big win. And the second big win with the integration of AI that we have had within my company is the AI marketplace. So what we have offered is especially in the field of pathology, we cannot force a particular AI tool to be used by doctors. We should give them the flexibility as to the kind of AI tool that they want to run. And since there are so many AI tools coming up in the field of cancer diagnosis that, we should make our platform flexible to accommodate those multiple AI tools. So what we have created on our platform is this AI marketplace wherein. Multiple AI companies all over the world could come onto our platform as buttons, and we make it quickly accessible to the doctor and research community and say they want to run tumor detection on breast from one AI company, and they want to run, tumor detection on lung from another AI company. And they want to compare the result. They want to understand how a tumor looks on breast versus a lung. Now they could easily do that on the platform. So that's another workflow adoption that we have seen. And the, the pharmaceutical companies and biotech companies are loving that idea. So giving, giving the flexibility in their hands.

Brooke:

That's fantastic. So it sounds as a scientist it's very beneficial, but what about when it comes down to actually tangibly saving lives and preventing cancer? How is AI supporting that?

Akash Parvatikar:

Yeah. Well, I think that's a billion dollar question that you just asked Brooke. But then Again, simplifying things here. So I am in the field of accurate cancer diagnosis but not in the treatment side of things. So all, my efforts and energy goes into the solving the problem that say a patient comes to a clinic, gets their biopsy done we want to able to at least diagnose that particular case correctly because then, the treatment sort of gets affected by the diagnosis that happens. So yes AI is playing a large role in drug discovery and, getting these new therapeutic drugs and all that into the market. But there is a very, very big problem at hand here, which is. Even for breast cancer, close to a million women in US undergo breast biopsy every year.

Brooke:

Mm-hmm.

Akash Parvatikar:

It so happens that majority of the cases are pre-invasive. That means they, they haven't had got a breast cancer yet, but it has a high recurrence of becoming a breast cancer in the next 10 years or 15 years. So now this pre-invasive is a notorious stage of breast cancer wherein there is a lot of disagreement among the doctors. Whether to call it a stage one or a stage two should they go for a surgery or, just lifestyle changes and all that. So, so now all of my work is focused on using AI to bring doctors into consensus, having concordance among the doctors to call out for a diagnosis. And AI is definitely helping that. And one more thing is that not just US, but all over the world, we are facing a shortage of pathologists. So what that means is now pathologists are signing out more cases than they can handle. So that increases their burden and that could also cause burnout and internally to misdiagnoses. So AI is definitely helping in automating some of the routine tasks, which for them might have taken hours to do. So that's where we are getting at.

Brooke:

Thank you for breaking that down further for

Akash Parvatikar:

yeah.

Brooke:

Getting more into the mindset of the future. What do you hope to see AI makes possible in the field of medicine in the next five to 10 years?

Akash Parvatikar:

Couple things here, Brooke. So one thing is we'll be seeing a much quicker and accurate diagnosis. Not just for cancer, but it could be other disease types which involves x-ray scanning or CT scanning, being able to detect features within a CT scan, x-ray scan and all that, that will all be automated, digitized, and, it'll be much more quicker. And that's one thing which is an achievable win with the help of AI. Second thing is in the field of drug discovery, where they trial out so many combinations of drugs before it comes into the market. And some of it was taking a lot of time just because the computation power was not there. And with the help of LLMs right now, they can pass through millions of data points and, try to find that one compound much more quicker. So again. It's about speed that I'm talking about for diagnosis and treatment. But third very interesting thing is that what AI could help is making us understand the field of medicine much better than we already did. For example, if it is cancer diagnosis, we with our naked eyes, might not have seen these features in the past, which the computer can now easily pick up. So this will lead to updating the medical records and medical books with the help of AI that, I think looking at this feature makes more sense than what we paid attention to early on. So, that's more of a question mark,

Brooke:

Mm-hmm.

Akash Parvatikar:

exciting, open-ended but very exciting time to be in. So.

Brooke:

Very exciting indeed. My next question for you is to share about any challenges you're facing or have faced with using AI. Have there been any moments where things didn't go to plan or

Akash Parvatikar:

Yeah.

Brooke:

I think that's important to touch on because

Akash Parvatikar:

Yeah.

Brooke:

think that AI is just the light switch, green go answer for them, but you know it's a process. So if you could share maybe about those, any initial lessons?

Akash Parvatikar:

Yeah. Again a great question that you asked Brooke. I wanna keep it simple and whenever we talk about AI, I think we need to talk about data. Because there is no data, there is no AI. So, I think to reframe your question whether I have faced the challenges with AI in my experience I have faced challenges with the data that I deal with. These are pathology dataset. And to give a context of the kind of data that I deal with whenever a biopsy is done by a physician and when they take a digital image of that tissue slide at 20 x or 40 x resolution. So if there was a giant printer that you might have and you had to print out this entire image on that printer, every single image occupies anywhere between two to three tennis courts. So that is the volume of that's one image and every single image, the size of it is anywhere between 500 megabytes to several gigabytes. So compared to an image that you might take on your phone, it might be, I don't know, say 10 megabytes or something. Every single image in on with pathology is few hundred, megabytes towards two gigabytes. So I had to face a lot of challenge in trying to get the data in the right format to not just build AI, but run any AI system on top of that, having to chop up the images into multiple square patches and then cook stitch it back, and making sure that, your system is quite powerful to handle this large amounts of data. So I think that was one challenge that I faced on, which took me a lot of time. Even more so than building the AI system because, and this is an open problem, not just with my research Google is trying to solve it. Microsoft is trying to solve it. Now Google has their own digital pathology wing. Again, a lot of funding. And Nvidia is getting into this field as well, wherein they're building special systems to handle this large volumes of data. So, different people might be in different field, but I think having this fair understanding of data and the challenges that come with it is very crucial to know very early on. And the second challenge for me is any AI that I build I sort of talk to domain experts very early on. I don't delay that because what I feel is that at the end of the day, they are the ones who will be using this AI system. So you cannot build your AI, have a product in place, and then do these customers for beta testing and all that. I feel that, if you have to beta test a product in the space of AI you have to recruit domain experts early on. For example, if you're building a AI, I don't know, virtual clothing app, right? You have to talk to fashion designers very early on. How do they think of fashion? How do they think of these different style so you have to sort of have this communication between the AI and the domain expert. So that's one thing that I learned during my experience.

Brooke:

One quick question because

Akash Parvatikar:

Yeah,

Brooke:

sparked in me as

Akash Parvatikar:

sure.

Brooke:

How do you prevent hallucination because you shared that you are dealing. With tennis courts of data, how are you preventing hallucination?

Akash Parvatikar:

So, one thing is, I do not deal with LLMs for pathology images, but there are companies out there, very interesting companies who have built LLMs for pathology images and I'm more than happy to leave a link to that to this episode in the description. So mine is more computer vision algorithms. For me using AI to identify features within the image and trying to understand the importance of those features in for a particular diagnosis.

Brooke:

Thank you for

Akash Parvatikar:

Yeah.

Brooke:

In your personal experience, are physicians responding to the growth and evolution of AI? It's starting to creep in, and completely rework the way that we experience healthcare. Maybe you can, at a very high level, what you've seen amongst your peers. Is it excitement? Is it fear? I'm just curious to hear what is the general response to AI and its advancements in the medical industry?

Akash Parvatikar:

Narrowing down to the field of pathology because medical industry is so, so wide.

Brooke:

Yes.

Akash Parvatikar:

so talking about pathology, Brooke there are current practitioners who have just graduated from med school and some early pathologists who are quickly adopting digital solutions. But there still are experienced pathologists with over 35 years of experience wherein they are still hesitating to go digital because they're completely wired in to look at the slides under the microscope, zoom in, zoom out pan and all that, and now they're a bit hesitant because for them they are set on that and they trust their decision making skills. So for them it's okay, why should I look at a digital file? I have been doing this for decades. So there I think we are having a very hard time going digital because there's a learning curve for them, but they are at the stage of their career wherein they are they have mastered the field and so that comes back to the hospital where if some of the pathologists are not willing to go digital then the hospital might take a back seat. We are trying to convince to go digital to make these AI systems work. It is two different challenges. One challenge is first produce the data from physical to digital, and then convincing what AI solutions will actually help them. The industry is trying to figure out baby steps into this. But to give our viewers a good information on this aspect, Mayo Clinic have scanned all their physical slides. This is over 12 million slides. NYU is scanning all their physical slides now, so they're going entire digital. The pathology department, U Miami is doing a great job and we have been talking to U Miami as well. And they're going very big in digitizing their physical repository as well. Hospitals are seeing value. Mount Sinai is doing it. Sloan Kettering is doing it. But these are all large hospitals who are doing it and I guess once they set a trend, maybe other hospitals and biotechs will follow. So.

Brooke:

Have you already seen the misdiagnosis rate go down quite dramatically?

Akash Parvatikar:

Dramatically, no,

Brooke:

No.

Akash Parvatikar:

But, but we have conducted experiments wherein, the rate of disagreement is going down. So again, what I mean by that is going digital. What's happening is say I'm in Texas. As a patient, but I have this pathologist who is in New York City and say, Miami. Now by going digital, I can easily transfer my data to these two doctors in, in seconds and they can look at the same image at the same time and give their opinion second opinions. So that's game changing. And what's game changing is that now the rural patients are getting access to the best care. From the best in class pathology. So they do not have to travel to the rural areas. They can sit in the comforts of their house and still diagnose the patients. All that they need is a scanner to digitize the flights and, get it to them. So it's called telepathology, which is, being able to remotely diagnose this patient. So that's happening. And there there hasn't been a clean report out talking about how much a discordance has come down or, misdiagnosis has come down. But if there's something that comes up, I'm more than happy to share that with you.

Brooke:

Thank you for that insight. That already sounds like such a advantage and I love that you highlighted that as well. In addition to time saved, it's the access to this care which I think must be very rewarding to be able to be a doctor and be able to support patients and make a difference in someone's life every day. What's one key takeaway that you want listeners to have when it comes to integrating AI?

Akash Parvatikar:

So I have two answers to that. Brooke. One is if you are interested to develop your own AI. The one key takeaway is try to understand the data, spend some time, spend some quality time in understanding the nuances of data that you are dealing with. It could be in any field that you are in even if you are not aware of the domain sit with KOLs, key opinion leaders and domain experts. What are the challenges that they are facing because think of AI as a solution to a problem. And I think problem comes first, solution comes next. So don't use AI to create new problems and then find solutions. Try to find existing problems, and then use AI to find quicker solutions. So my key takeaway I would say meditate over the problem for some time before you develop a solution. And again also spend some time in the kind of AI you are building with, because right now with agentic AI, LLMs and all that, that things happen so quickly wherein you might miss the train on some very key aspects and you never know, why the AI might be hallucinating in some cases. It's finding the right solution for everything uh, LLMs might not be the answer. if it's low hanging fruit, you can build a different solutions very easily. And that's something what I do I also act as an AI advisor to some of these startup companies wherein I just advise them how do you think about an AI solution? For the problem that you have and actually evaluate for yourself, using the ai, is it sort of improving your work efficiency?

Brooke:

Thank you for that final point there, and to your point, I will absolutely link whatever it is that you wanna share in the show notes, but my final closing question for you is. How can listeners reach out to you and connect and learn more about your work?

Akash Parvatikar:

They can connect me on LinkedIn. They can reach out to me via email as well. I'm easily accessible. I am in New York, so if any of the viewers are in New York, New Jersey happy to grab a coffee as well. And if any of you are in the stage wherein you have an idea and you want to develop an AI based product, and if you want a quick maybe 30 minute chat on, what it makes the best sense what kind of AI to implement and all that. Happy to chat about that as well. Because I have spent years with my research in building the AI system that you don't have to spend those many hours. And I, I'm more than happy to collaborate with, with our other products as well. So.

Brooke:

That's so generous. Thank you, and I really appreciate you taking the time to chat about a topic a lot of people are very interested in learning about AI in this intersection and specifically within your niche of pathology, so appreciate the time that you took to speak to me today. Thank you.

Akash Parvatikar:

Well, thank you. Thank you so much, Brooke. It was a pleasure talking to you. And a lot of interesting questions. It gives me a high level perspective of where I am in the field that I'm in and, because again, whatever AI we build we are answerable to the general public, not just the expert, so the kind of question that you ask, thank you so much for that. And I'm sure many of them out there who might not be aware that, there is a lot of good promise out there, especially in the field of cancer diagnosis and treatment. And there is definitely a big hope.

Brooke:

Wonderful. So exciting to hear.

Akash Parvatikar:

Yeah.

Brooke:

you.

Akash Parvatikar:

Thank you so much.

Brooke:

Wow I hope today's episode opened your mind to what's possible with AI. Do you have a cool use case on how you're using AI and wanna share it? DM me. I'd love to hear more and feature you on my next podcast. Until next time, here's to working smarter, not harder. See you on the next episode of How I AI. This episode was made possible in partnership with the Collective AI, a community designed to help entrepreneurs, creators, and professionals seamlessly integrate AI into their workflows. One of the biggest game changers in my own AI journey was joining this space. It's where I learned, connected and truly enhanced my understanding of what's possible with ai. And the best part, they offer multiple membership levels to meet you where you are. Whether you want to DIY, your AI learning or work with a personalized AI consultant for your business, The Collective has you covered. Learn more and sign up using my exclusive link in the show notes.

People on this episode