I am new to ML and looking to learn with some project. I have a medical imaging dataset where an image (image is a time series of an object so multiple images) has been looked at by radiologist and they have graded it on a scale of 1-5 for some pathology.

Now, I would like to basically use this to predict the pathology on new images. I am guessing there are multiple approaches one could take to do so. Could someone point me to some methods I could try (simple to more advanced) as I would like to also learn about them.

Another issue is that different images are of different sizes. Is this usually a problem for these approaches? I could try something where I can register them so that they are of the same size.

  • $\begingroup$ How many images do you have? That definitely constrains the kind of model you can apply. $\endgroup$ – Davidmh Oct 13 '16 at 14:23
  • $\begingroup$ Well, not too many. A couple of hundreds....but this can grow later. My whole idea is also to learn. So, I would like to get an idea of what are the intuitive approaches that more experienced people would take with this. $\endgroup$ – Luca Oct 13 '16 at 15:38

I have worked on similar projects (using medical images such as PET to predict outcomes). A method being used more and more for predicting cancer treatment outcomes is texture analysis: https://www.ncbi.nlm.nih.gov/pubmed/21321270

Another method of texture analysis uses wavelet transforms: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3505569/

After deriving the texture features, you could then use those in a ML model for making predictions.

A benefit to texture analysis is that it is mostly independent of the size of the target. It is however, more dependent on resolution. If your images are produced on different scanners, you may need to resample in order to normalize the images.

There are fewer studies about deep learning for medical imaging, but it definitely has some exciting potential. The question is whether or not a CNN can pick up subtle differences. From what I've seen (I'll look for a reference), CNN hasn't outperformed any other method yet.

Maybe add a few more details so we can get a better idea of what your objective is?

| improve this answer | |
  • $\begingroup$ This is very interesting. Thank you. Could you comment on how it is independent of the image size? In terms of feature engineering, I can use several features like SIFT or HoG (which come from computer vision). However, I am yet to imagine a way where this is independent of the image size. Also, different images may yield different number of these features and how do these models take care of that? $\endgroup$ – Luca Oct 13 '16 at 17:19

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.