Last month I presented a talk on my PhD project at the 2016 Aus and NZ College of Radiology and Radiation Oncology annual meeting, which is the first time I have talked about it publicly. We have not published any major papers yet (our first big one is just about ready for submission), so there hasn’t been much information on what we are doing.
I will explain a bit about the project in this blog post, but feel free to watch the talk. It only goes for about 15 minutes (around 10 minutes of talk plus a few good questions), and it gives a pretty good overview of the topic for anyone not familiar with machine learning and image analysis.
On a side note, my talk actually got tweeted/meme’d by Dr Andrew Dixon, who is among other things the Academic Director for Radiopaedia. Every sensible medico knows Radiopaedia the greatest radiology resource in the world, and vastly outstrips most textbooks for quality and breadth.
7 retweets. So this is what it feels like to be famous.
Without further ado – my research.
Precision radiology is, probably obviously, the application of radiology to precision medicine. Kind of like putting -omic on the end of any discipline (i.e., genomics, proteomics, radiomics) makes it sexier, so does putting “precision” at the front. I also get to use the term “artificial intelligence” to talk about my research, so I’ve got the whole trifecta.
These terms seem to frustrate people and cause much controversy, but I personally think both -omic and “precision” labels are quite useful because they identify a unique way to do science.
Precision medicine is the latest label for the application of -omic techniques to medicine. –Omics (roughly meaning “all of”) are the use of huge volumes of data with objective biomarkers to identify subtle patterns that human doctors can’t really find because we don’t see enough patients, the relationships are too complex for our poor meaty brains, and we can’t explore the parts of biology we don’t understand. More concretely, if genetics investigates single genes and their mutations to identify cause, effect, outcomes and so on, genomics is investigating all the genetic material and its variation in one go. Considering there are 3.2 billion base-pairs in the human genome, exploring mutations one at a time would take a while.
Obligatory double helix picture, but made of lots of dots. Because dots are data, and we have so much of it. Huge data.
I’ll talk more about hypothesis-driven research and hypothesis-free research some other time, but what it boils down to is that wide scale, big data, high-throughput research of the -omic sort is very efficient at finding correlations, but can’t really get at causation. The two methods go hand in hand, it isn’t a case of -omic science replacing the traditional scientific method (people did get upset about this in the past). I like to think of these techniques as an efficient screening system to identify the needles in the haystack, the unexpected places where further research will be useful. You want to search through a million pieces of data, you want -omics. You want to prove a certain drug targets a certain protein, you turn off the computer and hit the lab bench.
So precision medicine is the application of these techniques to medical practice. We use known, robust correlations found by -omic techniques to identify patient subgroups that matter in medical practice. The classic example is BRCA positive patients, a genetic mutation strongly associated with breast cancer. These patients are now offered preventative mastectomy because the lifetime risk of fatal cancer is unacceptably high, even with aggressive surveillance. Since the discovery of the BRCA mutations, many thousands of lives have been saved by this targeted therapy option.
And that is the point – precisely targeting therapy. That isn’t to say that we haven’t been targeting therapy in clinical medicine, in fact every test we do is for this purpose. Diagnosis itself is an example of this concept – a diagnosis identifies a patient subgroup that will have a certain natural history and respond to certain treatments. Unfortunately, diagnosis is pretty blunt, and two people with the same diagnosis may have wildly different histories and treatment responses. The key difference in precision medicine is that we use an -omic method to gather as much information (data) as possible to reach some theoretical limit of how precisely we can define patient subgroups. Hence the value of the terms ‘-omics’ and ‘precision’.
The same concept has been called personalised medicine, individualised medicine and many other things. It is no exaggeration to say that this is the big medical idea of our age, and our best (and possibly only) shot at beating the major diseases that cause ill-health and death in Australia and worldwide. Which is why it is so hot right now – it isn’t just smoke without fire, this seriously going to be the future of medicine.
Radiology has so far been left behind as many other parts of medicine have jumped on the precision bandwagon. It hasn’t been entirely clear how we radiologists can be precise in an -omic sense, with the few forays in radiogenomics gaining high visibility but so far minimal clinical impact. It may just be too soon for these techniques to translate into clinical practice, but I will leave the discussion about the strengths and weaknesses of current methods in radiomics and radiogenomics to another day.
A NEW ROLE FOR RADIOLOGY IN PRECISION MEDICINE
Precision medicine has a problem – most of the major successes have come from genomics. Don’t get me wrong, genomics is the greatest advance in medical science in recent history, but genomics has a major flaw when it comes to measuring human health. The genome isn’t actually responsible for most of our disease.
The majority of disease in the developed world (and increasingly in the developing world as well) is chronic disease, and the genome is only responsible for 20-30% of chronic disease risk. This is probably obvious to everyone – it is our lifestyles that lead to diabetes, emphysema, heart disease, and so on. Once we start looking at people over the age of about 60, chronic diseases start to account for something like 90% of morbidity and mortality, and a huge chunk of our healthcare costs. To throw in a few new words so I can sound very smart, the lifestyle component of health is called the exposome, which is the sum of lifetime environmental exposures. This includes things like diet, pollution, stress, and so on. The combination of the genome and exposome is called the phenotype, which describes the totality of variation in human health. Lifelong and lifestyle, nature and nurture.
So precision medicine has so far been unable to get at the bulk of disease that afflicts our populations because we don’t have a good way to assess the critical part of the phenotype. Many approaches have been attempted to quantify the changes of chronic disease – proteomics to measure circulating serum proteins, epigenomics to analyse changes to genomic expression and function post-conception, microbiomics to assess the commensal organisms that inhabit our bodies. And many, many, many more.
A small subset of -omic domains. I have no idea what orfeomics is.
But none of these techniques get at the underlying problem. Chronic disease is the build up of cellular and subcellular pathology or damage, such as protein aggregates in Alzheimer’s, cross links in arteriosclerosis, lipid-laden macrophages in heart disease. To measure chronic disease with any level of accuracy, we have to measure this damage.
Microscopy is the gold standard here, but we can’t go around subjecting everyone to dozens (hundreds?) of biopsies to get an accurate estimate of the burden of disease. Particularly because most organs (like the brain) don’t really like biopsies, and because a tiny sample of tissue via a biopsy needle isn’t actually very accurate for this sort of thing.
Cross-sectional imaging (CT, MRI) can visualise all of the tissue at once. It can identify and quantify the pre-clinical changes of chronic disease, like with coronary artery calcium scores. There is even evidence that sub-visual image features (things computers can see but humans cannot) correlate to cellular and molecular variations such as tumour subtype and genetic mutations. These were the hints that led me to my research, because these properties of images sound very much like the building blocks of a precision medicine test for chronic disease. If only we could find the important features in the images.
We need an -omic solution to images. Current radiomics methods go part of the way there, but so far haven’t really explored what we might call the “space of all possible image biomarkers” very thoroughly. If genomics is the science of measuring every change that can possibly occur in the genome, we need to work out what the imaging equivalent of a base-pair is. What are the fundamental building blocks of medical images, and how can we efficiently analyse the entirety of this mammoth domain? Again, a complex discussion for another day.
But this is what we are trying to do. We are using more traditional radiomics methods alongside “deep learning” to try and find out everything we can about human health variation from medical images. So far we have been able to identify image features that account for around 70% of a persons risk of death within 5 years, using CT images alone (with strong controls for highly predictive clinical confounders like age). This is preliminary work, but is very exciting for us because it hints that our approach is more than just theoretically attractive. We are about to scale up from our proof of concept experiments to a massive dataset, from tens of patients of tens of thousands. Because big data is better. Seriously, there are mathematical proofs of this sort of thing.
I will try to keep this blog moving along, with more explanation of the various concepts I have briefly touched on here. I might even weigh in on the question I get asked most – are radiologists going to be replaced by computers in the near future (spoiler: sort of).
And never fear, I will try to explain deep learning as well, because, well … it’s so hot right now.