Geoffrey Hinton believes we should stop training radiologists, and Andrew Ng claims that there are already machine learning models that perform better than radiologists. Should radiologists be worried? Should we soon expect radiologists on the side of the road holding signs that say "Will read x-rays for food"? This talk analyzes the evidence underlying these claims to separate out the hype.
The gauntlet has been thrown with respect to computer vision and radiology. Leaders in the field of machine learning and artificial intelligence have decried the end of the radiologist is in sight, while radiologists counter that they are irreplaceable. In this talk, we'll explore the current state of the field when it comes to automated diagnosis in radiology using machine learning.
To get a better view of what we can accomplish today and what is still missing we will start by comparing the tasks performed by the algorithms to the job required of a clinical radiologist on a daily basis. We'll then examine the unique difficulties inherent in applying machine learning to not only radiology, but any healthcare dataset, and why we can't just simply replicate the process used in other computer vision success stories, like self-driving cars. The last part of the talk will question the task itself, i.e., is automated diagnosis of radiological images the best use of our time and resources, or are there other more fruitful avenues to explore.
This talk is for anyone interested in the intersection of machine learning and healthcare and is suited for all levels. There will be some discussion of different neural network architectures, but only at a high level.