Better machine learning methods for medical imaging segmentation

Question

I am wondering if there are some better machine learning methods for medical imaging segmentation?

Currently, I have some relatively low-resolution MRI images and I tried to use histogram and k-means algorithm, which is an old method to do a general clustering based segmentation. However, pixel may help estimate more than one cluster in the k-means procedure.

Is there any novel direction for clustering pixel values in order to do the segmentation?

I think this can be done by two ways:

1. Change the clustering similarity measurements but still use k-means, e.g., change histogram based clustering into something like using mutual information, but it might be well done by other people (if you know any, please give me the reference).

2. Change k-means to other clustering methods but keep the similarity measurements, i.e., histogram.

Any ideas would be appreciated. Thanks a lot.

Answer 1

One of the most popular methods for doing image segmentation is Markov Random Fields: essentially, this forms clusters of pixels, and this is moderated by a connectedness prior that says neighbouring pixels like to be in the same cluster. Pushmeet Kohli's work is a good place to start, as he's demonstrated how graph-cut algorithms can find the MAP solution exactly.