Paper: Boundary loss for highly unbalanced segmentation : https://arxiv.org/abs/1812.07032
This is a demonstration of Boundary Loss for refining segmentation of fine and rapidy changing boundaries on an inhouse dataset. We tested for patches and whole images. We tested this on various medical datasets and usually got an improvement of 1-2 dice scores or of around 3%. We first trained our model to saturation on dice and cross entropy and then continued training by adding this loss. We dynamically changing the weightage in the loss according to the paper but not as severely. We found lower learning rates and momentum to be helpful. Over all we found this a great loss using level sets to focus on boundries. All files to be uploaded soon.