Assessing the Promise of AI in Oncology: Potential Role in Cancer Screening, Diagnosis, Staging

EP: 1.Assessing the Promise of AI in Oncology: Definitions and Goals

EP: 2.Assessing the Promise of AI in Oncology: Potential for Precision Medicine

EP: 3.Assessing the Promise of AI in Oncology: A New Journal

EP: 4.Assessing the Promise of AI in Oncology: A Diverse Editorial Board

EP: 5.Assessing the Promise of AI in Oncology: Call for Evidence

Now Viewing

EP: 6.Assessing the Promise of AI in Oncology: Potential Role in Cancer Screening, Diagnosis, Staging

EP: 7.Assessing the Promise of AI in Oncology: Improving Patient Interactions and Experience

EP: 8.Assessing the Promise of AI in Oncology: Matching Patients With Clinical Trials

EP: 9.Assessing the Promise of AI in Oncology: Looking to the Future

In this sixth episode of OncChats: Assessing the Promise of AI in Oncology, Toufic A. Kachaamy, MD, and Douglas Flora, MD, LSSBB, FACCC, discuss potential opportunities to leverage artificial intelligence (AI) tools in cancer screening, diagnosis, staging, and prognosis.

Kachaamy: I think this is very applicable, as you mentioned, [with regard to] cancer screening. We definitely need a better cancer screening process in this country. How do you see AI changing cancer screening? Where do you see cancer screening 5 years from now, using [these tools]?

Flora: I think that that’s probably where we have the most success today. These tools are ready. There are dozens of companies that have validated instruments and they’re putting a lot of resources into this, whether it be Google or Philips or other [companies], they can help us with risk stratification and review things that the human eye could not discern in three dimensions. Risk stratification [allows us to know] who is at risk for developing cancer later. Obviously, we’ve talked a lot about image analysis, whether that be augmenting the read of mammography in those very dense breasts with lots of cysts in young women, or looking at really indistinct lung nodules on a sagittal cut of a lung cancer screening CT.

We’ve talked about pathology, I think there’s some exciting stuff that we could talk about [like] getting into liquid biopsy, circulating tumor cells, and those sorts of [approaches]. And then finally, I think there’s a role for big data to scrape the electronic medical record and say, “Hey, listen. This patient who had a biopsy by a dermatologist out here in this surrounding community suggests a possible genomic disease and can build tools that will get that patient referred for genetic screening or germline screening, whereas the primary care or the dermatologist didn’t maybe recognize that fact.” So, lots of promise there in [terms of] screening.

Kachaamy: So, what about [with] cancer diagnosis? What can you tell us [about] where AI is now in terms of cancer diagnosis?

Flora: Yeah, well, we’re getting there. There are 3 or 4 papers that have been published in the past 2 months [that are] looking at improving diagnostics, both from MRI and CT mammograms. Google DeepMind has a system that can look at lung cancer CT scans with greater accuracy and positive predictive value than dedicated chest CT radiologists. I think you and I are probably both seeing a lot of changes in this liquid biopsy world where we’re not having to go back and re-biopsy and re-biopsy for pathology reports to guide us in terms of these next-generation sequencing [approaches]. There’s work being done by Exact Sciences in GRAIL and Freenome and that area. There are lots of areas where I think there’s a lot of work to be done but I would say cancer diagnosis and screening are probably [areas where] we can expect [to see more] this year or next.

Kachaamy: Perfect. So that [will] really impact people soon in terms of practice and clinical outcomes [for patients]. Do you see anything coming down the pipeline about cancer staging and prognosis?

Flora: I really hope so. I am a survivor of cancer twice now. I had a resected kidney cancer 6 years ago, I should be okay, and I currently have prostate cancer that’s been gauged by DeCipher, which is an AI-led tool [that looks] at RNA expression of 22 different genes and [uses] AI to sort risk categories that are probably more useful than traditional staging. We think so. Now, it’s being used as a way to augment our ability to predict those things. However, I think the ability of us to not just look at shape, size, and lymph nodes, has been evolving.

[For example,] ulcerated melanoma gets upstaged. It’s not a big stretch for me to think that the American Joint Committee on Cancer v. 10 might include the presence of circulating tumor DNA for [patients with] resected colon cancer to help oncologists make adjuvant treatment decisions. Those data are already in a very early stage of publication. But we think there’s some utility and some predictive benefit for tests like Signatera to look at colon cancer circulating tumor cells. So, I do hope and expect that we’re going to be using these genomic signatures and these risk categorizations to add to stage grouping with more selectivity, rather than just [looking at] how big and how wide it is. I think we have smarter tools today, as soon as they’re validated.

Kachaamy: So, moving closer to precision oncology.

Flora: Absolutely.

Kachaamy: Now, I’ve seen some research on predicting cancer genomics from histology, which can give you faster answers you can use immediately at the time of diagnosis. Do you have any information on that? What can you share with us on this? Do you [think this is] promising?

Flora: I do. It feels like a step back from genomics, but it’s not actually that. When we move to routine histology, there’s a gigantic potential cost savings, and when we talk about democratizing things like screening tests, you can’t do a $2,500 MRI of the breast on every woman in America for breast cancer screening. Similarly, H&E digitalized slides, where they take thousands of pixels and look at this with pattern recognition software, may take us forward in a way that is much less expensive and ready for primetime sooner without that gigantic dataset required for genomic assays. I think we’re going to see more of that, and there are a number of companies that I’m aware of that are doing that right now, trying to derive genomic information from routine histology. It’s a very clever way to use data we already have.

Check back on Monday for next episode in the series.