Machine learning for medical imaging analysis
Everyone talks about machine learning. Why is it not already commonly used for medical imaging analysis? What are the challenges and blockers?
Computers are already great at chess, Go, and even some complex computer games. Especially Deepmind, a subsidiary of Alphabet is quite advanced. Also it can nicely detect whether there’s cats in videos.
Algorithms, processing power, and data are the ingredients for machine learning. One exception are situations with very clear rules like chess. Deepmind’s AlphaZero for example was training solely based on the rules of chess. Though for more complex, real-world problems that usually doesn’t apply, which means that you need data for training of your machine learning system.
What is machine learning?
From an application perspective, machine learning is split in two steps. The first is the training. And the second is the classification. The data that you usually need for machine learning are needed for training the system and validating it. The trained machine learning system is then fed with new data and gives you its output based on the training data. That means that a machine learning system can only be as good as its training data. And that also explains why machine learning systems can be racist. If you only use images of white middle-aged men to classify humans, then the system will most likely only classify white middle-aged men as humans. It’s not the machine learning system that’s racially biased, but the person building the set of training data.
What data do I need?
Regarding the kind of training data, there’s one big question. Are they labeled or not? Labeling means tagging each image with the respective classification of that particular image. A simple example would be using two different classes for classifying images. The one label is “cat”, the other label is “no cat”. And each image either has a cat or no cat.
Unsupervised learning is not as straightforward to apply as there’s no guidance for the system what it should detect.
Supervised systems are the much more systems. But as you can already imagine, besides having good data, you also need good labels. If a lot of the input data is wrongly classified, then the machine learning system can’t be trained properly.
Traditional machine learning vs new and fancy deep learning
Traditional machine learning systems consist of two stages. The first stage is about transforming the input data into simplified data that the machine learning system can work with. This is usually called feature extraction. If you have images of sticks and you want to classify it by the material, some of the features that you define could be the color, the length, or the thickness. You need to define anything that you assume will help the actual machine learning system to make sense of the input. And those features, in this case the color, length, and thickness, is being used as input for the machine learning system (e.g. a neural net).
And then, as described above, the system will use those features together with the labels, e.g. wood and metal in this case, for each individual image and get trained based on it. If you then have new images of sticks you can feed it into it. First it calculates the features and then classifies it with the trained machine learning system.
That’s at least how machine learning worked for a long time. Nowadays you hear a lot about deep learning. And the great aspect about them is that you don’t need this feature extraction step. It’s done within the deep learning system itself and doesn’t need human guidance. That’s obviously amazing because it’s not limited what humans consider useful for the classification, but can leverage any information for which the computer with all its processing power finds any correlation to the label information.
So, what’s the downside of deep learning? You need a lot of data for the training, much more than for traditional machine learning approaches.
Long story short, you want a lot of high-quality labeled data to be able to use deep learning, which would make you benefit most from machine learning.
Why should we use machine learning in medical imaging analysis at all?
Let’s look at how human doctors learn how to analyse medical images. Quite simplified, they look at a lot of images and learn how to classify them and what to look out for. Well, that’s not so different from how a machine learning system works. Though a computer can process many more data than a human can. A machine learning system has quite some advantages:
- can use all existing training data from all over the world
- can be updated regularly
- doesn’t get tired
- it can get replicated easily
- a computer can process so many more data in parallel (e.g. x-ray, MRI, US, PET)
Even if doctors would never get tired, they still can only digest a specific amount of data for learning. And what would you prefer as a patient, a doctor that was trained on a subset of medical images that were available to him or a system that was trained with all medical images that are available worldwide? Also maybe you have a rare disease that your specific doctor may not have seen before. He will most likely not recognize it.
And combining several data sources, maybe for a full diagnosis a combination of your PET and MRI scan are necessary. Combining multiple, potentially multi-dimensional sources of information is pretty hard for humans.
Human doctors are still very much needed and ideally work together with a tool as described above. Human doctors can verify the results of the computer and are also much better in human communication.
Again, why is this not done yet?
As discussed a lot of data are necessary for this. And there are not many databases of medical data or images available. This is one of the reasons for Deepmind working the the British NHS, as they are a nationwide health system that can give access to huge amounts of data.
There’s a lot of great researchers out there, also processing power is not the limit. The main bottleneck to make progress on the topic of machine based analysis of medical images is to have sufficient data available.
Photo by Harlie Raethel on Unsplash