How can I mark a superpixel that belongs to a region of interest?

Question

I've applied superpixels to segment an image (more precisely graph-based watershed), and it has 1000 segments (i.e. 1000 superpixels). A bare example is shown below:

The image I'm working on is a teeth x-ray (grayscale). I want to select all the superpixels that belong to the teeth and then train my ConvNet for classifying them as "teeth" or "non-teeth" tissues.

After I classify them, I would like to get all the ones with "teeth" labels and mark them over the original image until a tooth is totally segmented. An example is shown below, where the referred tooth is marked in red.

So how can I do this automatically, that is how can I classify the superpixels as "teeth" or "non-teeth" tissues?