Hey guys. Someone posted this thread in hacker news, and I joined this forum to offer some clarification. I don't think
anyone going into radiology or medicine need worry about their careers rights now. Medicine is a highly regulated field and change of this magnitude is going to take decades (I think). Healthcare deals with people's lives, so naturally there will be a lot of push back and distrust against automation like this. As someone who has built machine learning systems for finance companies, it's difficult to convince executives why we should follow my blackbox model against all the industry acquired domain specific knowledge acquired over the decades. The only thing we have for us is the statistics, the results of the validation on true real life data. But it's hard to convince people with just the cold hard statistics and numbers. This is going to be 10x worse for something like medicine, because people's lives depend on it, and we
should be highly skeptical.
With that said, it is now a matter of time until machine vision starts dominating the field
. It's an eventuality, and here's
why.
1. In traditional machine learning,
experts (doctors for example) work with computer scientists to hand engineer "features" to solve very specific problems. For example:
http://ieeexplore.ieee.org/abstract/document/1617200/
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4485641/
These traditional techniques rely on domain specific expertise to engineer features. Once the right set of features is selected, simple linear models like naive bayes or K-means outperform human experts (real life physicians).
The issue is that feature engineering is really hard, expensive, tedious, cumbersome and time consuming. Most importantly, feature selection is
specific to each problem. Every time you want to solve a different problem, you have to re-do feature selection.
2. Deep learning is gaining so much traction because it
automates feature selection. Rather than manually engineering the features,
unsupervised learning techniques and deep architectures allow the problem to be expressed in terms of feature layers that encode higher and higher level features.
Finally we have a general technique that can be used not only on different problems in the same domain, but this same technique can be used
across entire domains. As it is, deep learning is being utilized in all sorts of disciplines including speech recognition, not just machine vision.
Someone here said that the problems radiologists are solving are very different from being able to recognize cats in videos and what imagenet is solving. This is not strictly true. It turns out, we can train a deep belief network with a bunch of random images, and the network still learns useful features/abstractions such as being able to recognize edges, corners, contours, blobs, etc.
3. Recent advances in the field allowing the use of GPU instead of the CPU for training, convolutional neural networks, dropout, generative adverserial networks, transfer learning amongst other factors has caused a lot of positive movement in the right direction specifically in computer vision.
Here's an amazing overview of the potential of deep learning in Radiology:
http://www.cs.ucf.edu/~bagci/publications/DLradiologyRSNA2016.pdf
Here's the type of efficiency a machine vision system can provide in pea sorting. Each pea is analyzed, and air jets are used to knock the bad ones away. This is pretty common already:
https://arxiv.org/pdf/1602.03409.pdf
"We study two specific computer aided detection (CADe) problems, namely thoraco-abdominal lymph node (LN) detection and interstitial lung disease (ILD) classification. We achieve the state-of-the-art performance on the mediastinal LN detection, with 85% sensitivity at 3 false positive per patient, and report the first five-fold cross-validation classification results on predicting axial CT slices with ILD categories. Our extensive empirical evaluation, CNN model analysis and valuable insights can be extended to the design of high performance CAD systems for other medical imaging tasks."
One other important take away from the paper above:
However, there exists no large-scale annotated medical image dataset comparable to ImageNet, as data acquisition is difficult, and quality annotation is costly.
This is currently a huge (largely artificial) limitation. The paper discusses a couple techniques to deal with the lack of data, notably transfer learning, that allows unsupervised pre-training on already existing large natural image datasets like ImageNet. Once we start getting quality, annotated/labeled medical images, I'd expect to see progress skyrocket in this area
Finally, they achieved the 86% state of the art result by pre-training on ImageNet. Training the convolutional neural network on just the medical images from scratch yielded 70%, which is still better than the traditional approach of expert hand selected features.