This is The Age of AI Series, where we talk to the foremost entrepreneurs and innovators around the planet using ML to transform industries. (Join our special mailing list!)
Recently, there’s been a lot of interest in using AI to help doctors diagnose cancer. But we don’t often hear about using AI for the treatment of cancer, which is what we’ll talk about today!
Radiotherapy is a type of cancer treatment where you burn tumor cells with very precise rays. The majority of cancer patients get prescribed radiation, and have been for 2 decades.
My guest is Mahmudul Hasan, the Founder and CEO of MVision, which builds AI software for radiotherapy treatment planning.
- 01:02 — How radiation therapy works
- 02:05 — The “positioning problem” in radiation oncology
- 08:47 — Why it’s critical to understand the different phases of diagnosis and treatment
- 12:23 — Shortening the time-to-treatment for cancer patients using AI
- 16:08 — The problems caused when doctors are inconsistent
- 19:47 — Challenges of building a “clinical-grade” ML product
- 30:14 — Designing a world-class engineering process for clinical-grade ML
- 36:55 — Dealing with crushing regulations as a medical tech company
- 42:53 — MVision’s team and growth
Aman’s 2-Minute Summary and Key Takeaways
Why AI in radiotherapy
Radiation therapy is the classic form of “precision medicine,” which has become a buzzword recently in the medical-tech field. Our story here begins after a patient has been diagnosed with cancer, and the doctors have prescribed radiotherapy as a treatment.
The next step in radiotherapy is to create a proper treatment plan where the doctors have to exactly determine the tissue area that will be targeted, as well as the right dosage of radiation that won’t harm healthy cells.
Traditionally, this was done by doctors looking at 2D scans of the area with their naked eye, and scribbling some notes. The more modern hospitals have software where they do the “segmentation” digitally with a mouse or stylus. But since it is a very manual process, and requires a LOT of training and experience, there’s a lot of inconsistency — the same doctor will annotate the same scan differently each time.
This creates operational inefficiencies due to which the treatment plan sometimes takes up to 2 weeks to get prepared.
MVision builds AI software to automatically segment the cancerous tissue from healthy cells, and streamline the treatment planning workflow. As a result, certain hospitals have even been able to conduct same-day treatments!
How do you decide if an AI software is “clinical-grade”, and can be used by doctors? Turns out, it’s not an entirely objective criteria — it’s a classic case of “I know it when I see it.” The doctors see the AI-generated contours, and see how well they correspond with the underlying anatomy of the organ they’re looking at. If they deem that it’s accurate (or at least better than they themselves are), then it’s clinical-grade.
Another point that came up is building an AI team and an engineering process for achieving clinical-grade accuracy. The last mile is always the hardest, thus the team invested heavily in training data and also a “golden dataset” of human-labeled scans.
Mahmud and his team seem to have achieved phenomenal growth, because they worked closely with their end users (doctors and oncology departments) to figure out the pain points, and built a solution that makes their jobs demonstrably easier, which helps with adoption.
And for anyone interested, they’re currently raising a growth fund.
(Ethics Policy: These opinions are 100% my own as an independent observer and educator. I don’t own stock in guests’ companies or their competitors, nor do I get paid by them in any form for any reason at the time of publishing, unless specifically stated. Episodes are also not intended to be an automatic endorsement of any company or its products and services.)