What I learned in RSNA Radiology in the Age of AI Spotlight Course
How will artificial intelligence effect Radiology?
Introduction
About a month ago I had a chance to attend the RSNA Spotlight Course “Radiology in the Age of AI”, where even Dr. Ng gave a talk on his take on the subject. (And I actually had a chance to take a photo with him! :D).
Here I will summarize some of the materials that I have learned for each session, unfortunately, I was not able to attend every talk, but I hope some of the material that I’ll discuss can be a help to you as well.
AI is the New Electricity: The Disruptive Power of AI Applications in Medical Imaging — Andrew Y. Ng, PhD
Dr. Ng had a great influence from his father. His father created a machine learning algorithm (way back in the day) that aided doctors.
Most of the benefits of machine learning these days come from supervised learning settings. And thanks to the abundance of data AI can now automate a lot of different tasks.
This is great an all, however, without a large amount of data today’s learning method would not have been successful. Especially in the medical industry where we don’t have a lot of data.
A model that works well in the research lab might not work well in the real world since the data distribution is different. This change in distribution is very problematic.
So the question is, how to keep the generalization power of the model? There might not be a clear answer yet, but best practices need to be built in radiology, as a form of regulation.
Additionally, AI will be used to augment Clinician’s ability rather than replacing them. Computer scientist and radiologist have to work together.
Fireside Chat and Question & Answer
To educate both radiologist and computer scientists making a course or a Bootcamp is a good idea. The key to innovation is bringing people together, to solve some of the serious issues we are facing today in Radiology.
One of that issue is Data labeling, and how to extract information from radiology reports. (More related to NLP tasks).
Breaking Down the Mystery of the Black Box in AI (XAI) — Safwan Halabi, MD
How can we explain the model’s decision? Most of the methods that we have are strong black box models (as well as weak black box ones as well).
Neural Networks are hard to understand and audit, due to having so many parameters. We have no idea what individual role a neuron plays in the system’s point of view, and how each connection is made.
Another way of putting this issue is the problem of dimensionality since we have so many parameters it is beyond human capability to understand the model with that much parameters. Explaining ‘what caused what’ is a hard challenge to tackle.
Finally, there is a transparency problem. (what part of the input data is the model looking at?). Sure we have methods such as deconvolution, integral gradient, Deep LIFT and more, but can we say that this problem has been solved?
What Radiologists Need to Know about Data Science — Hugh Harvey, MBBS BSc
As a radiologist (or researcher) if you want to use Data science (machine learning), should you build your own algorithm or should you collaborate with a company? How clean are the data?
Those are some interesting ideas to think about, also they have developed a level of data cleanliness ranging from level D to A.
Level D → Raw data
Level C → de-identification of the data, ethical approval is needed to access the data. Finally, it is unstructured.
Level B → Structured Data and quality control, no labels
Level A → Cleaned and Labeled data
From all of those steps, getting the ethical approval of the data takes the most amount of time.
Bias and Implications for Medical AI — Matthew P. Lungren, MD, MPH
Started the talk with an urban legend of using a neural network to classify tanks, like that the model we use might have a bias.
When computer systems have an unexpected outcome, we need to know where those errors were made and why they were made.
In the medical field, there can be three types of bias hospital, Computational, and Cognitive.
Hospital bias → different hospitals use different machines to capture data in different manners. A model that has been trained from one hospital data source might not do well from another data source.
Computational bias → Adversary attacks.
Cognitive bias → this one is more of human bias since humans tend to behave differently in different situations.
AI in Breast Imaging — Hugh Harvey, MBBS BSc
Interestingly a lot of data related to breast cancer are originating from Europe.
Some problems in this line of research include different companies using different machines and pre-processing methods to create the mammography images, how can we standardize them?
This different sources can create bias also patch based analysis was not successful for breast cancer mammography images. Even for segmentation, data labeling is key. Since segmentation is the backbone of a lot of different processes if we don’t get this right might cause some problems.
AI in Body Imaging — Bhavik N. Patel, MD, MBA
CT images are extensively used in medical settings, and there is a lot of potentials when paired with AI, starting from segmentation and classification.
However, there are also a lot of challenges, in diagnosis such as lung or prostate cancer classification, noise to signal ratio is very critical. And overcoming the serious imbalance data distribution is another challenge. (paired with a small number of training data it gets harder).
Even though the model might have been trained on natural images, using pre-trained models are a very good idea.
Final Words
It was an amazing conference, and I got to meet Dr. Ng! That was a life-changing moment for me. Please click here, to see program details and more. I hope some of the information here helped you as well.
