Hcfyk

I have the opportunity of doing a PhD under the supervision of an expert in medical imaging at a top institution. Currently their group does not conduct research into the application of machine learning to medical image acquisition and processing. The purpose of the PhD studentship would be to pursue research into this. The department has significant machine learning and signal processing research groups whose seminars I will be able to attend and academics I can have contact with.

The supervisor has not for some time (before deep learning) pursued research in machine learning. The PhD itself is as yet not strongly structured and will initially require a deal of exploration and prospecting before its final form is decided.

Given that there is a safe fallback of medical imaging I do not foresee a risk to completing the PhD. However, as the only member of the group pursuing machine learning would this be a very risky PhD to embark on, particularly considering that afterwards I intend to pursue a career in academia? Are there any benefits?

I would ask about having a co-supervisor. Having access to esteemed DL researchers is great -- but they will have limited time/interest in helping you if you are not "formally" their student. If you manage to find someone in this role, I think your position is just about perfect.

If you don't manage to find someone in this role, I have three main concerns:

• You will spend a ton of time re-inventing the wheel. For example, can you train a CNN on ImageNet from scratch? There are a lot of caveats needed to obtain state-of-the-art results (e.g., dataset augmentation, regularization loss, etc.), and you will likely rediscover them one-by-one (or, use a black-box model you don't really understand). A DL expert would likely already have working code and could explain it to you, allowing you to jump right to the research. (Yes, there are open source codes...but in my experience, they all require a lot of work to be both transparent and accurate.


• Mathematical rigor. It's easy to just learn ML/DL at a "technician level" -- but as a PhD in it, you should really understand it a mathematical level if not a theorem/proof level. It can be difficult to do this on your own.


• Problem selection. Your medical advisor will likely find it super novel to run existing techniques on medical images. There may even be a novel application here, on the medical side -- but on the ML side, this is not really interesting, it's just a straightforward application of one technique to a straightforward problem. This is maybe OK if your interest is entirely on the medical side -- but if you want to do something also interesting on the ML side, you would essentially be on your own to come up with something. That will be difficult to do (for the first time) without advisors on both sides.

Those are the main blind alleys I see. Of course, there is also a ton of upside -- this sounds like a very interesting, prestigious position that would position you well for an academic career. Only you can judge this tradeoff.

Do you want to design a tool that can build many things, or learn how best to use the available tools to build a house?

Do you want to do a PhD in machine learning or are you trying to use machine learning to solve problems in medical imaging?

In the first case I would agree with @cag51. Without a Deep Learning supervisor, it would be challenging and also unlikely your PhD would reach its full potential.

However, if you are more interested in finding novel and practical uses for existing machine learning techniques in order to improve the field of medical imaging then the lack of specialist supervisor is less important. There is a startling amount of low hanging fruit which requires only a broad conceptual understanding of machine learning combined with domain-specific expertise (e.g medical imaging).

After your first paper/project you will no doubt discover a host of problems that are specific to your domain area which require further research and in-depth knowledge of the domain area which can be provided by your primary supervisor.

It could be a great opportunity to help the field take advantage of benefits provided by machine learning in a very applied and practical way as well as carve out your own niche in academia.

Sounds like a great fit, with some options for different paths post-Ph.D. along with some fallback if things don't work out perfectly. I wouldn't be super concerned about having all kinds of supervision by a deep expert. It is common for grad students to do their own work without significant apprenticeship by the "advisor" (grant writer). As long as you are careful to look out for yourself by sticking to tractable problem(s), it should be fine.

In addition, you seem to have thought things out and expressed them well. And some of your comments (like department work in signal processing) show enough awareness that you seem to be able to look out for yourself and drive your own research.

I agree with Jonno Bourne's answer, but I don't have enough reputation to comment.

I just want to add that I was in this same situation during my PhD. Specifically, I was in the second scenario, so if this is what you pursue, I can say from my experience that it is perfectly viable. You will just have to learn a lot of stuff on your own, but this is the cool part of a PhD, isn't it?

If instead you want to do a PhD on machine learning, as opposed as using machine learning, then I would too consider looking for a (co-)supervisor with ML expertise.

Yes, this can be possible to do. I would not consider it particularly risky.

One of my professors when I was a MSc student did almost exactly this when he did his PhD once upon a time. He specialized in one learning method and built applications for it in his main supervisors field.

But it was long time before "deep learning" existed and subsequent ML-trends appeared. So I imagine it should hardly be more difficult now to motivate than it was then.

The idea of trying to get a co-supervisor with good skills in learning seems like very good advice.

I recently graduated with a PhD in Plant Breeding. At my university, an increasing number of students are working with building predictive machines that their major advisers have no experience with. Most of us (myself included) were first students in plant breeding that applied predictive machines in largely technical manner, producing fairly derivative research from a predictive perspective, but was novel based on the crop is was applied to. The students that excelled in this situation the best were those with significant modelling experience to begin with, and almost all had completed Masters. If you go this route, you'll need to be more self-directed than average, and prepare to teach your major advisor as much as they teach you. I struggled a lot with the lack of clear direction from my major advisor, but it was ultimately worth it, as it opened up more options than would have been available if I took a more traditional path.

By undertaking a PhD in a field you are pursuing an interest in a field you just don't want to know more about, but you want to become an expert in. Your supervisor should be someone who can effectively guide you through this field and, when needed, teach you.

PhDs are difficult and to give yourself the chance of learning the most possible I would embed yourself within a ML research group so you can learn from the best, rather than stumbling through the field yourself.

While co-supervision is an option, my experience of it is it often does not work, with the student feeling stranded between two supervisors who each mutually see the student as the others problem. Great supervisors could work in synergy, but unless you have a way of evaluating this before you start you would be taking a gamble.

From the supervisors perspective, what they need is to collaborate with an already established ML research group or bring in a Postdoctoral researcher with a PhD in a relevant ML sub-field. I think this is more likely to be a success and less risky for everyone involved.

On the matter of funding: Unless you are extremely wealthy don't consider doing a PhD without funding. There are many PhDs out there with funding and not enough good people to do them...

Closing statement:

If you consider yourself an ML expert and want to learn more about medical imaging, do the PhD with the mentioned supervisor. If you want to become an expert in ML and maximise your chances of success in an academic career then undertake a PhD with the best ML research group you can get a funded PhD with. Save the cross-field collaboration for your post-doc or latter end of your PhD.

No matter what decision you make, doing a PhD is a great privilege, so make the most of it and don't look back!

Important questions to answer for yourself:
1. Jonno Bourne's question of what area do you want to focus on is important. In other words start with the end in mind.
2. Most of what you learn will not be from your prospective supervisor and his recent experience with ML is not very important. Are you someone comfortable defining your own path?
3. What do other graduate students working with your prospective supervisor think of him?
This is important. He may want a particular outcome and will limit your investigations or he may encourage creativity and let you decide how you can contribute.

Funding for my RA was very important and it gave me peace of mind.