AI4BH – COMP0172
Coursework #2



Geoffrey Hinton, one of the most prominent computer scientists and deep learning expert
said in 2016: “We should stop training radiologists now. It's just completely obvious that
within five years, deep learning is going to do better than radiologists”.

He further argued that “Any old problem where you have to predict something and you have
a lot of data, deep learning is probably going to make it work better than the existing
techniques”. We are in 2022 and there is a lot of evidence of the potential of deep learning.
In this coursework you are asked to comment on G. Hinton’s statement. Should we stop
training radiologists?

More specifically, you are asked to write a 2000 - 2500 words essay structured as follows:


Context (10%): What is radiology? What kind of techniques are used in radiology? What
does a radiologist do?

Rationale (10%): What are some of the current challenges (2 at least) in radiology that
would benefit from machine learning?

Argumentation (20%):
Pro: Can you give some arguments in favour of G.H. statement? (2 arguments)
Cons: What about against? (2 arguments)

Andrew Ng, another pioneer of modern machine learning said: “I want to live in an AI-
powered society. When anyone goes to see a doctor, I want AI to help that doctor provide
higher quality and lower cost medical service.”

Comment (10%):
What does A.N. mean by that? How can AI increase quality of care while lowering costs in
radiology?

Robustness (20%): What are the current challenges in deploying deep learning solutions for
radiology in clinical practice (3 challenges at least)?
How would you tackle those challenges? (Unlimited data is not an option).

Fairness (10%): Explain how machine learning solutions for radiology can be unfair? Why
would that happen? How would you stop it from happening? (More data is not an option)

Ethics (10%): What ethical consideration should be considered?
Example: what are the potential dangers of private companies having access to radiology
data? What about AI models able to identify sensitive attributes from radiology images?

Overall presentation (10%)

References:

You need to defend your statements with references and, if required, you need to describe
the technical solutions. Below are a few papers related to the topic, but you are more than
welcome to find your own or use others such as the suggested readings from the lectures.

NON-EXHAUSTIVE reference list:

Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in
under-served patient populations
https://www.nature.com/articles/s41591-021-01595-0

Gender imbalance in medical imaging datasets produces biased classifiers for computer-
aided diagnosis
https://www.pnas.org/doi/abs/10.1073/pnas.1919012117

Rajpurkar, P. et al. CheXNet: radiologist-level pneumonia detection on chest X-rays with deep
learning. Preprint at https://arxiv.org/abs/1711.05225 (2017).

Radiologist shortage leaves patient care at risk, warns royal college
https://www.proquest.com/openview/ff4e0a3098380a9f39485f360c19bebc/1?pq-
origsite=gscholar&cbl=2043523

Reading Race: AI Recognises Patient's Racial Identity In Medical Images
https://arxiv.org/abs/2107.10356

Robustness: https://spectrum.ieee.org/andrew-ng-xrays-the-ai-hype (really interesting to
read/watch)