Save a lot of money and save a lot of lives — an opportunity for Australia.

We are at a point now with some areas of medical diagnosis where not much money can be invested for significant savings both in healthcare expenditure as well as lives.

Much has been written about how AI/Machine Learning has a long way to go in terms of replacing the general intelligence as expressed by normal humans . However what this thread masks is a number of areas where software has overtaken human ability and with the exponential characteristics of technology the software gets even better very, very quickly.

For the most part the places where software overtakes human ability is where the boundary conditions are very clear (or pretty clear) and the data is clear of significant levels of noise (errors). Games is an area where even going back to Big Blue beating Kasparov in Chess in the late 1700’s (ok it felt like that long ago!) technology has had the upper hand.

The best example I can think of in this space is the Asian strategy game of Go. The first data point is that in the early 2000s the best AI folk believed that Go would not be “solved” by software for another 50–60 years..

Then in 2015 the brilliant people at Deepmind produced AlphaGo which had been trained on 100,000 games of Go played by humans. This software version of Go took on the world champion and beat him 4–1 and then cleaned his face on the second set.

Probably most interesting (and less written up) is what happened late in 2017. In Q4 2017 Deepmind released a report on AlphaGo Zero which was a new version of the software. This time the software was provided with the rules of Go and nothing else! No games data, strategy hints etc..The result? AlphaGo Zero played itself 1m times over 3 days and then took on AlphaGo..It beat AlphaGo 100–0.

In areas where the rules or decision processes are pretty clear software can learn to make better decisions than humans.

The first area outside of games that we will see this expressed is in the Health area. So much of disease diagnosis is really just pattern recognition — and in the case of pathology, dermatology etc it is just image recognition.

Globally there are a wide variety of projects underway where well structured, sufficient data sets are being used by Data Science teams to come up with better ways to diagnose disease. Sadly a lot of the data sets that exist are full of noise (errors..where for example a particular MRI is tagged with a diagnosis that ends up being wrong…not very helpful).

However notwithstanding this issue Deepmind is doing amazing work both with the NHS (see this link to get a taste of the project set Deepmind is focused on in health

Google is also (noting Deepmind was acquired by Google but is not completely integrated as yet) doing other cool projects. One example where they are achieving significant traction is in the area of Diabetic Retinopathy ( where it seems that the image recognition software is way better than a panel of specialist eye doctors.

Also the UK Government has decided to focus on this area as well (

Of course disease diagnosis is not the final answer. In the case of Diabetic Retinopathy once the disease has been diagnosed some very smart and experience specialists need to work out the specific therapies for the patient to help them not lose their sight. We are perhaps moving from a time, with SOME disease conditions (where the data has low levels of noise), where diagnosis will be the realm of software but treatment selection and delivery will remain (for a while at least) the realm of well trained specialists.

Where should Australia play in all of this? Australia has 2 advantages relative to our small scale where we can do something globally interesting. Firstly there are very few countries that have massive, all of population medical records that cover all treatments and all drug therapies supplied. The UK has it with the NHS and we have it with Medicare/PBS.

This is an advantage because AI/Machine Learning software needs small teams of Data Scientists to help develop the inference sets but they HAVE to have access to well tagged (low noise) data sets that the software can use to learn.

Australia also has a medical research community that in scale and quality if well beyond what our population size would suggest.

Perhaps given the above we should try to focus on areas where our data sets and medical research capability gives us a leg up — perhaps areas where Australia is over-represented in terms of disease conditions.

One clear winner is Melanoma.

Australia and New Zealand have melanoma rates higher than anywhere else in the world and compared to the US our relative rate of melanoma deaths is nearly 3 X that of the US (

At this point let me declare a self-interest which I do not think clouds my view but rather informs it.

3 years ago my middle daughter was diagnosed with Melanoma. Fortunately we had access to amazing dermatologists (shout out to Dr Pascale Guitera) and surgeons (shout out to Dr Jonathan Stretch). The good news is the cancer had not spread and our daughter is now in a high rotation regime of checks.

As I watched our daughter go through the process of diagnosis and compared it to what I went through 30 years ago when I had a Basel Cell Carcinoma cut off my leg it was clear that at its core initial melanoma diagnosis relied on someone looking carefully through a high resolution device to see if mole looked funky compared to other moles they had seen in their careers..If it looked funky then it was monitored and if still funky then it comes off and goes to pathology.

Melanoma is a terrible disease in that if you catch it early the cure rates are very high but catch it just a little bit later and your prognosis is awful.

Stage 0 melanoma are really just abnormal cells sitting on the top layer of skin and have yet to grow down into the epidermis. The 5 year survival rate for Stage 0 is 100%.

Stage 1 Melanoma. In stage 1A, the tumor is up to 1 millimeter (mm) thick. It also has no ulceration, which means the tumor hasn’t broken through the skin. Stage 1B can means two things: the tumor is up to 1 mm thick and has some ulceration, or it’s between 1 mm and 2 mm thick and has no ulceration.

The five-year survival rate for stage 1A is 97 percent and 92 percent for stage 1B. The 10-year survival rates are 95 percent for stage 1A and 86 percent for stage 1B.

Stage 2 Melanoma. Stage 2 melanoma means the tumor has grown more than 2 mm thick. Doctors will also analyze the tumor to see if it’s ulcerated. Surgery to remove the cancerous tumor is the usual treatment strategy. A doctor may also order a sentinel lymph node biopsy to determine the cancer’s progression.

The five-year survival rate for stage 2A is 81 percent and 70 percent for stage 2B. The 10-year survival rates are 67 percent for stage 2A and 57 percent for stage 2B.

Stage 3 Melanoma. At this point, the tumor can be any size or shape. To be considered stage 3 melanoma, the cancer has to have spread to the lymph system. Surgery to remove cancerous tissue and lymph nodes is possible. Radiation therapy and treatment with other powerful medications are also common stage 3 treatments.

The five-year survival rate for stage 3 melanoma ranges from 40 to 78 percent. The 10-year survival rate ranges from 24 to 68 percent.

Stage 4 Melanoma. Stage 4 melanoma means the cancer has spread to other parts of the body, such as the lungs, brain, or other organs and tissue. It may have also spread to lymph nodes that are a good distance from the original tumor. As such, stage 4 melanoma is often hard to cure with current treatments.

The five-year survival rate is only about 15 to 20 percent. The 10-year survival rate is 10 to 15 percent.

So between the Melanoma being 1mm thick and > 2–3mm thick your 10 year survival rate goes from around 95% to around 15%. Pretty shitty odds. Which is why Melanoma is not fooled around with and why potential Melanomas are treated with speed and action.

So assume a dermatologist (or GP) sees a mole that they think is suspicious. If they do not detect it as melanoma (false negative) then in a matter of months, untreated, that patient goes from healthy to terminal (ok there are some amazing immunotherapy drugs like Keytruda for late stage melanoma but they only work in around 40% of cases so the promise of universal immunotherapy is not around the corner).

If the doctor is worried then they order a biopsy. Current protocols call for monitoring of moles that look suspect and then order a biopsy. It is worth noting that new technology like dermoscopes help increase the correct diagnosis rate by about 20%. So we have better surveillance (if you happen to be very close to, or have access to, great Dermatologists). Even then if you have the biopsy and this ends up being benign (false positive) then objectively speaking the patient has gone through some level of unnecessary pain and disfigurement and a bunch of health dollars were wasted.

It is worth noting that to get to a level where you are best in class as a Dermatologist you probably need to have seen close to 10,000 lesions ( I didn’t make that number up but rather got it from one of the leading dermatologists in Australia) — to allow you to develop the fine nuanced ability to detect the good from the bad. If you are a normal GP then your chances of seeing this many lesions is near zero and so your chances of being great at detecting melanoma is low.

Not surprisingly great dermatologists biopsy 5 moles with 1 being a melanoma (1 in 5 hit rate) with Australia GPs getting a hit rate of about 1 in 10. So for most of the Australian population there are between 5 and 10 biopsies to find one melanoma. A lot of false positives.

The full data on both false negatives and false positives is sketchy at best. A few studies (one here suggest that between 40% — 90% of skin biopsies turn out to be not required. There is no data on people who were let go when they actually had melanoma but other studies on the accuracy of visual decision processes would suggest that this is not an insignificant number.

In a better world we find a way to minimise false negatives and false positives while also ensuring the same level of diagnosis is available to someone who lives next door to the leading Dermatologists in Australia and someone who lives in a country town.

There was a fantastic paper in February 2017 Nature magazine where the issue of skin cancers and image recognition was referenced with regard to s/w trained to look for funky moles. The BEST summary of this paper (shout out to Tom Spacek) was written up by Dr Luke Oakden-Rayner who is a PhD candidate in the area of medical AI/Image Recognition.

For those that can’t be bothered to read Luke’s summary the highlight is that “…they achieved better performance than most of the individual dermatologists, as well as the “average” dermatologist, from their panel that they presented for comparison” ..Here is his writeup. It is worth a read..

So we know that a fundamental breakthrough in scaling up the use of AI/ML image recognition in the area of skin cancer is not some way out dream but within touching distance.

So what?

What if we could use this technology to easily lower the costs to the health system dramatically?

What if we could better decide who to do biopsies on?

What if we could use the technology to actually save lives?

The Cancer Council of Australia ( suggests that 750,000 Australians a year have treatment for non-melanoma skin cancers and around 14,000 for melanoma..Simple maths would suggest that if there are 14,000 melanoma treatments annually then there must be between 70,000–140,000 biopsies taken annually (1 to 5 or 1 to 10).

Depending on the consultation fee etc and which pathology lab was used the total cost (system cost) for a biopsy is between $350 — $1,000. So Australians are spending between $30m — $140m on biopsies a year. Of which at best 20% of the expenditure was necessary.

Of course while a biopsy has a direct cost of between $350 — $1,000 the total cost should take into account time off work to have the procedure and any time off work post plus also some way to cost the pain and disfigurement of patients.

The cost to families for false negatives is much higher. Again using simple maths the Cancer Council reports that around 2,500 people die each year from Melanoma. Given that we know if Melanoma is caught at Stage 0 then the 10 year survival rate is near 100% it is clear that (probably) 2,000 lives a year at least could be saved by making sure every mole could be accurately detected and tagged.

If you step back from lives saved, addressing false negatives also has the potential to save hundred of millions of dollars in treatment costs for people who get diagnosed with stage 2–4 melanoma and take a while to die. With 14,000 new melanoma cases a year starting their treatment journey current estimates indicate in Australia over 100,000 people are at some point in their melanoma treatment journey.

At one end there is the person who had a melanoma removed and now have 3 monthly checkups (annual cost around $500). At the other end people who are on expensive immunotherapy drugs after having been through many expensive surgeries and chemotherapy regimes (thousands of dollars a year. So to suggest by diagnosing early we could save $100m or more in treatment costs does not seem adventurous.

Of course sadly for many on the treatment treadmill this is not the end. Multiple surgeries and drug treatment, alot of pain and further disfigurement and then you die and all because we didn’t correctly diagnose the stupid mole in the first place.

So by just ensuring more accurate diagnosis we can save perhaps $100m in unnecessary biopsies by overcoming false positive ratios and then maybe hundreds more millions of dollars and 2,000+ lives more on the false negative side.

It is perhaps not surprising that Australia is a world leader in rates of Melanoma — a by product of living in a land with great weather! It is therefore also not surprising to note that Australia is a world leader in skin cancer research generally. We have outstanding researchers as mentioned above.

It seems that Australia is the perfect place to establish the world’s most effective AI/Machine learning skin cancer image recognition capability.

We know what the benefits of such a service would be (see above!). So what about the costs.

Well to create something meaningful we would need access to large data sets of tagged moles (probably >100,000). This data is in hand as groups such as The Melanoma Institute Australia are working with others (ISIC — International Skin Imaging Collaboration to aggregate data sets.

We also need a team of dermatologist researchers and a team of data scientists to collaborate to build this application layer on top of one of the global image recognition platforms (Google is probably the leader at this stage). We do not need dozens of people. Maybe a handful of researchers and a handful of data scientists.

And we probably need to find maybe $5m- $7m a year to run this project until it can be self-funding through commercialisation globally of its offering…So let’s say $30m over 5 years to be conservative. We keep the focus just on the issue of better use of image recognition and machine learning s/w to increase the accuracy of (and access to!) melanoma diagnosis.

What does the return on investment look like. We invest $30m over 5 years and on completion we maybe have a product that saves hundreds of millions of dollars in unnecessary biopsies and then (on the other side) expensive treatment for people diagnosed with melanoma and maybe 2,000 lives a year for ever!!! …Seems like a no brainer to me.

My family is prepared to invest $2m in this project. Anyone else interested?