Improve features function

[OlehKSS]

For classifier is will be better to train on the regions extracted by our algorithm, so we will need to change the way we find true positive of our segmentation function. See #68.

Use Jackard index or Dice similarity to find correspondence with the regions on groundtruth and on the segmented image.

We'll need to define decision threshold, which is around 0.2. It can be provided as a parameter for the feature function as well.