The repo is aimed for segmentation of overlapping chromosomes, as described in the problem statement given on AI ON website
The repo uses U-Net, state-of-the art segmentation net for segmenting overlapping chromosomes. The repo used Lasagne, a Theano based library for segmentation.
The data consists of 4 classes, where class 4 is the common region between 2 overlapping chromosomes. The classes 1 & 2 , are non-overlapping part of each of the chromosomes. Class 0 is the background
The performance of the net was observed using mean_dice_score. It was computed as
dice_score = 2*I/(GT + PL) where I is the sum of the number of pixels predicted correctly except background, GT is the number of pixels which belong to ground-truth except background and PL is the number of pixels in predicted image except background.
There were 2 methods of training attempted
- Treating all the classes independently (param combine_label = True in segmentation.py)
- Treating Class 1 & Class 2 as same i.e. Class 1 and Class 3 as Class 2 (param combine_label = False in segmentation.py)Assumption being the non-overlapping parts inherently aren't different in each chromosomes. Then we can apply conventional CV methods like watershed algorithm to distinguish between the non-overlapping blobs
- Build the Watershed algorithm post-processing part to the pipleline in case of combined_label training
- Optimise the parameters better for the nets in each case
- Build a full pipeline for test_data generator using best validated model
With combined labels, could reach a dice score as high as 0.97. Without combined labels, could reach a dice score as high as 0.81.