looking for radiology programs with computer science tracks

[hot_thicc_potato]

I'm a rising 4th year interested in radiology. I'm trying to compile a list of radiology programs with tracks in computer science or AI. Here's an example from University of Washington. Any help would be much appreciated!

---

Members don't see this ad.

 

[stcl]

Washington is great as you mentioned.

Ohio State has a very impressive AI development group where they've worked with NVIDIA; you can do either a mini-fellowship or individual research electives. Prevedello is also well-known in the machine learning field. He was leading the RSNA panel on machine learning when I went.

UC Irvine started up an AI in medicine program in the past couple years. And I don't know particulars about Stanford's programs, but I know that several of Stanford's radiology faculty (such as Langlotz) are well-published in the radiology AI space.

I'm sure there's other programs like this, but I'm not familiar with every program.

 

[808s&heartbreak]

I don't get it. If you want to learn deep learning, then learn deep learning. Why do you need a "track" to do it? As long as you get your clinical duties done and are capable of being a competent clinical radiologist, nothing stops you from learning deep learning on your own. You don't need a track. Often times these tracks funnel clueless radiology residents into doing scut work (annotating data sets).

 

[hot_thicc_potato]

I don't get it. If you want to learn deep learning, then learn deep learning. Why do you need a "track" to do it? As long as you get your clinical duties done and are capable of being a competent clinical radiologist, nothing stops you from learning deep learning on your own. You don't need a track. Often times these tracks funnel clueless radiology residents into doing scut work (annotating data sets).

I appreciate your observations; I thought it might be easier to find mentors in the subject if the program gave it formal attention

 

[Cognovi]

I appreciate your observations; I thought it might be easier to find mentors in the subject if the program gave it formal attention

It would be more worthwhile to compile a list of faculty who have presented or published something of interest.

 

[deleted564680]

CCF has an IT track/pathway but I agree with above...if you're interested in it then just do it. Some institutions have more resources than others though.

 

[Naijaba]

I don't get it. If you want to learn deep learning, then learn deep learning. Why do you need a "track" to do it? As long as you get your clinical duties done and are capable of being a competent clinical radiologist, nothing stops you from learning deep learning on your own. You don't need a track. Often times these tracks funnel clueless radiology residents into doing scut work (annotating data sets).

Hi, I haven't posted on this site in a while. I am one of the current instructors of the Deep Learning pathway at UW.

It is possible to learn the terminology of machine learning using available web resources. However, to actually understand machine learning, you have to train / test / deploy your own models. Unfortunately, there are many hair-pulling hurdles that stop people before they even get off the ground. Most of these aren't related to machine learning and are idiosyncrasies of computer science and mathematics. For example:
-Do you know what a Terminal/Console/Shell is? What about the commands bash, cd, ls, tar, zip, chmod?
-What is the difference between a png, jpeg, jpeg-2000, DICOM?
-How do you translate the image resolution into millimeters using the DICOM header information?
-How many bits in a byte? In a word? What's an integer versus a float? A double?
-How many bytes does it take to store a 512x512 color image? What about a single 512x512 grayscale CT slice? (Hint: It's more than the color image).
-What is a DLL? What does Linux/Mac OS call these? (Hint: blue screens of death are almost always related to DLLs)
-How do you multiply a matrix times a vector? Inner/outer/convolutional product?
-Do you have CUDA installed? How about CuDNN? Do you have the libraries installed needed to even compile these on your system?

If these sound painful to you, they are! All of these are essential skills to learn, and we only scratch the surface during the deep learning class. Computer science, mathematics and machine learning have their own vocabularies that are rarely encountered in clinical medicine. This is what makes learning machine learning difficult to do on your own.

Conversely, one of the best benefits of our hands-on Deep Learning course is having people around you who know what they're doing. This helps you avoid getting stuck in a "PC LOAD LETTER" scenario. If you've ever tried machine learning eventually you will encounter a bizarre error such as: i) two copies of a linear algebra library, neither of which work (because her friend used it to do some research years ago and other libraries fell out of sync), ii) an error related to the Windows PATH environmental variable (no spaces in the pathname, please), or iii) permission errors (gotta give that Python notebook somewhere to save its progress).

The goal of the class is to make residents comfortable with the foreign vocabulary of machine learning. You will be far from an expert in computer science or mathematics, but you should be able to read ML journal articles, and train your own models.

 

[808s&heartbreak]

Hi, I haven't posted on this site in a while. I am one of the current instructors of the Deep Learning pathway at UW.

It is possible to learn the terminology of machine learning using available web resources. However, to actually understand machine learning, you have to train / test / deploy your own models. Unfortunately, there are many hair-pulling hurdles that stop people before they even get off the ground. Most of these aren't related to machine learning and are idiosyncrasies of computer science and mathematics. For example:
-Do you know what a Terminal/Console/Shell is? What about the commands bash, cd, ls, tar, zip, chmod?
-What is the difference between a png, jpeg, jpeg-2000, DICOM?
-How do you translate the image resolution into millimeters using the DICOM header information?
-How many bits in a byte? In a word? What's an integer versus a float? A double?
-How many bytes does it take to store a 512x512 color image? What about a single 512x512 grayscale CT slice? (Hint: It's more than the color image).
-What is a DLL? What does Linux/Mac OS call these? (Hint: blue screens of death are almost always related to DLLs)
-How do you multiply a matrix times a vector? Inner/outer/convolutional product?
-Do you have CUDA installed? How about CuDNN? Do you have the libraries installed needed to even compile these on your system?

If these sound painful to you, they are! All of these are essential skills to learn, and we only scratch the surface during the deep learning class. Computer science, mathematics and machine learning have their own vocabularies that are rarely encountered in clinical medicine. This is what makes learning machine learning difficult to do on your own.

Conversely, one of the best benefits of our hands-on Deep Learning course is having people around you who know what they're doing. This helps you avoid getting stuck in a "PC LOAD LETTER" scenario. If you've ever tried machine learning eventually you will encounter a bizarre error such as: i) two copies of a linear algebra library, neither of which work (because her friend used it to do some research years ago and other libraries fell out of sync), ii) an error related to the Windows PATH environmental variable (no spaces in the pathname, please), or iii) permission errors (gotta give that Python notebook somewhere to save its progress).

The goal of the class is to make residents comfortable with the foreign vocabulary of machine learning. You will be far from an expert in computer science or mathematics, but you should be able to read ML journal articles, and train your own models.

I think it's a good idea but it is unclear what kind of radiology residents these pathways target, and what the goal is.

For residents without any technical background: It would be near-impossible for someone in radiology residency to go from not know what a bit is at the start of residency to learning the stats, multivariate calculus, matrix algebra and concepts of object-oriented programming, and learning a programming language, and then be able to fully deploy a deep learning model - all while working full time and studying as a radiology resident.

For those with technical background: these people already know the answers to the above questions (or can google the ones they don't know and understand it without difficulty), can troubleshoot errors, and install CUDA, know unix shell commands, rudimentary python and OOP. To train your own models, you have to get your hands dirty and actually code/fail/reiterate, which requires time and consistency. In these cases, the limiting factor is time during residency, not the lack of a "deep learning" pathway. I speak as someone who fall in this pathway.


What you described sounds good in theory but I am confused as to what the end goal of these pathways are. In practice I think it will produce half baked radiology radiologists and a half-baked deep learning practitioners who are not competent in either.

 

[Cognovi]

I think it's a good idea but it is unclear what kind of radiology residents these pathways target, and what the goal is.

For residents without any technical background: It would be near-impossible for someone in radiology residency to go from not know what a bit is at the start of residency to learning the stats, multivariate calculus, matrix algebra and concepts of object-oriented programming, and learning a programming language, and then be able to fully deploy a deep learning model - all while working full time and studying as a radiology resident.

For those with technical background: these people already know the answers to the above questions (or can google the ones they don't know and understand it without difficulty), can troubleshoot errors, and install CUDA, know unix shell commands, rudimentary python and OOP. To train your own models, you have to get your hands dirty and actually code/fail/reiterate, which requires time and consistency. In these cases, the limiting factor is time during residency, not the lack of a "deep learning" pathway. I speak as someone who fall in this pathway.


What you described sounds good in theory but I am confused as to what the end goal of these pathways are. In practice I think it will produce half baked radiology radiologists and a half-baked deep learning practitioners who are not competent in either.

Maybe it's for those who fall in the middle: people who know what a bit is, knew a programming language in high school, know stats, have the patience to troubleshoot errors, and can google things.