There are a few things that are wrong.
First, screening mammography is extremely common so many women are undergoing screening mammography that will reveal some abnormality.
The way those abnormalities are next assessed is by doing biopsies.
About a million biopsies a year come back with non-invasive diagnoses, which could range from normal breast to something like atypical ductal hyperplasia, to ductal carcinoma in situ.
Each of these carries an increasing level of risk of either actually having cancer in the breast at the time or for developing future breast cancer.
So the problem is this is one of the diagnostically most challenging areas of pathology and several studies have shown significant discordance in diagnostic rates when people are given the same sample, that different people will do different things.
What is it about mammography that makes it so easy to not get the right answer and not be specific?
It’s a screening test so people really, on the one hand, don’t want to miss an invasive cancer so they might set a lower bar for calling something abnormal that should be evaluated further.
It’s totally non-invasive so the radiologist doesn’t have access to the actual cells to make a precise, specific diagnosis.
All they can say is they see some pattern that appears abnormal and a lot of benign things can produce abnormal patterns, like different benign neoplasms or different inflammatory conditions in that area.
Now, you are there, ready to rescue the situation with lots of computational power. What are you doing? What kind of factors can you put into the algorithm?
Once the tissue is removed and we have the microscopic slide, the same slide that the pathologist would look at, we can run computer algorithms over that to quantitate thousands of features characterising characteristics of the epithelial cells themselves which may go on to become cancer cells, the supporting stromal cells as well as relationships between all of these heterogeneous cell types.
Then after we generate these thousands of features from the image we can put that into a machine learning algorithm to find the combination of features that does the best at predicting a specific outcome, which could be what is the diagnosis or what is the patient’s prognosis.
That’s the physical appearance or the physics of the situation, what about biomarkers?
Yes, we can incorporate biomarkers into this overall platform, it would be one additional source of features.
So far, what have you done?
For what I’m going to be talking about at the meeting this week, what we’ve done is applying this to see if we could train the computer to discriminate with very high accuracy two of the lesions on the spectrum of pre-invasive breast lesions, going from one that’s considered completely benign, called usual ductal hyperplasia, to one that’s essentially treated like a malignancy, which is ductal carcinoma in situ.
We wanted to see if we provided the computer with expert diagnoses and with the images can we train the computer to discriminate these two automatically.
And how much of an improvement can you make?
What we showed in our study was in our first dataset we obtained an area under the curve of about 0.95, which is almost perfect performance in a totally automated fashion for separating these two lesions.
Then, when we applied this to a completely external dataset where the slides had been processed in a different way and was a very difficult validation set we obtained an area under the curve of about 0.86 for discriminating these two lesions using the exact same automated algorithm.
The concordance, though, needs to be exceptionally high to convince people because even with quite a small risk many women will still opt for an operation, won’t they?
Our goal is not, at least in the near term, to be replacing the pathologists; it’s really to provide… What I think would be very useful would be a real-time second opinion, in which case the computer can help spot the cases that the pathologist might have the wrong diagnosis on and could require further evaluation based on discordance between their diagnosis and the computer’s diagnosis.
What would you like cancer doctors to be doing about this sort of possibility now in their day-to-day practice?
I still think it’s a work… the technology is under development but what I hope is we’ll advance this technology to a point by applying it to large retrospective cohorts where we have a consensus diagnosis of very high quality as well as we have clinical follow-up on the patients where we can really build the robust, informative system that will outperform our competing practices today, at which point it would be great if a biopsy were taken and then we applied the system and then were able to return to the treating doctor highly reproducible, highly informative diagnoses that we’re