The nerve area called the brachial plexus needs to be segmented from the given images. This was a challenging problem since there were an uneven distribution of images with masks. Some of the kernels had pointed out data with contradictory masks.Therefore a simple thresholding could eliminate these images.
The image format was .tif and 580x420 in size.Therefore it is too big to be trained on my GPU. I have decided to rescale my images size 128x128 since it is faster to predict and train. I tried with all the training images initially for seeing what performance the data set without any augmentation provides. It gave a private score .53. The improvements occurs with augmentation and checking for those contradictory masks for the similar images.
This training set had many contradictory images . Histogram intensity can be found using
###spatial distance using cosine similarity
import scipy.spatial.distance as spdist
D = spdist.squareform(spdist.pdist(x, metric='cosine'))
A unet is always best for segmentation especially for a biomedical image analysis. I decided to use a Unet without pretrained weightsand used Dropouts in between to increase acuracy.The loss function used is a dice score which can be used for checking for overlap beween ground truth makss and predicted masks.
Currently trying to implemnt a FCNN model , will update soon