ֿDomain: Medicine / Oncology
Status reached: $20M term sheet
When: 2020, for 3-4 months
Team skillset: 2 second timers, one with CEO background and one with CTO background.
- Cancer diagnosis remains a largely manual process that involves a pathologist with years of training analyzing information.
- The idea was to use deep learning to analyze scans and identify things automatically and accurately.
- The outcome would be more accurate diagnoses and reduced manual labor.
Why it didn’t progress
We understood that on its own, the idea brought insufficient value, and that there were other companies doing similar things, including an Israeli company that Entrée invested in and a larger company in the US.
- Identifying new biomarkers using technology that doesn’t involve H&E staining of the sample.
- The problem it would solve: a better match between oncological patients and the medication they need. There are a lot of cancer medications and doctors can’t know which patient should get which medication. Today, tissues are stained with Hematoxylin and Eosin to identify the relevant biomarkers to try and figure out the best treatment.
- Research found that you can do the same via an analysis of the tissue using deep learning, which would allow doctors to identify biomarkers that don’t yet have a chemical process associated with them.
- Identifying new biomarkers would be an incredible breakthrough in medicine that would be worth a lot of money.
Why it didn’t progress
- Not enough scale, based on a few factors:
- The radiology market is 10X bigger than the pathology market and even companies in this world, like Zebra, are struggling. When you do the math, it’s good but not big enough money.
- Timing: maybe in 10 years when the world is more digitized this model could work.
- Industry potential: With the exception of Flatiron, very few startups in the medical field succeed.
- Realized not enough value in identifying a visua/scan
- This is just one piece of a long and complex process, which is mostly manual.
- The pathologist does a lot of things that aren’t just the identification of the biomarkers and the cancer. There is a lot of back and forth between the pathologist and the oncologist. There is a lot of information that is qualitative: the size of the cyst/growth and the type of cells, for example.
- In another example of the complexity, you have to slice the cells. If the slicing is done inaccurately, any analysis of the scan will be problematic.
- There are about an hour’s worth of tests that the pathologist needs to do, so a yes/no response on a scan does not offer enough value, since s/he will need to do a lot more work.
- This understanding was validated by the person who runs Siemens’ activity in Israel.
- Also spoke to a company called VIM.AI, which is trying to identify strokes. They have been very successful and think you need one of the following two things in order to be successful:
- Identify esoteric things that most doctors miss
- Change a hospital process. When dealing with strokes, time is a critical factor, so much of the solution’s messaging was around alerts for doctors, not just about the ability to identify the stroke.
A few challenges to consider when building a solution for this problem:
- FDA approvals: A system that is involved in a doctor’s decision-making process require FDA approvals and there is significant regulation.
- Lack of interest by pathologists: based on conversations with pathologists, there was little excitement about a solution, with reasons ranging from concerns about job security to this being a very conservative world.
- Deep learning: Deep learning has become commoditized, so it’s hard to prove that better deep learning leads to superior results and as a result technology is less of a factor in these companies.
- Working with hospitals: it is very difficult to change a process in a hospital, and scale is extremely difficult, because hospitals work differently from each other.
- US industry: the hospital industry in the US is very fragmented, making it difficult to reach scale since hospital networks/companies are made up of a very small number of facilities each.
- Data challenges: Data is inconsistent, tagged differently, and not very reliable.
- Analyzed companies: Zebra, Aidoc, ibex-ai, Nucleai and VIM.AI
- Consulted with a very senior, world renowned oncologist.
- Worked with two deep learning experts and an academic.
- Spoke with the person who runs the Siemens operations in Israel (Siemens creates a scan machine used in oncology).