C3W2A Assignment U-Net Segmentation Results

Hi everyone!, actually I don’t have any problems with completing the assignment but I didn’t really understand the training results. We want to distinguish between different tissue types but all the results are greyscale images that I can not comprehend how to read. What is the proper of reading those results? And if we wanted to colorize the outputs for different labels how can we implement it, what should be the number of output channels? I would appreciate any help :slight_smile:

Hello Meircy!
Hope you are doing well.
No, they are not trying to distinguish between different tissue types. The assignment is based on this paper where the goal is to convert EM (Electron Microscopy) image into an accurate boundary map, defined as a binary image in which “1” indicates a pixel inside a cell, and “0” indicates a pixel at a boundary between neurite cross-sections.
So, in the picture you shared, the first one is the EM image, the second is the ground truth label (Boundary map plotted by humans) and the third is our model’s output. So basically, the goal is to minimize the error between the 2nd image(labels) and 3rd image(predictions). So eventually you can see that the 2nd and 3rd images become similar as the training progresses.
If you want to colourize the output, you should have the ground truth labels in that form(Colourized) so that our model learns it.
Hope that this post will help you to understand. If not, feel free to post your queries.
Have a great day!

Oh I completely misunderstood what the assignment aims to do, now everything is clear. Thanks a lot Nithin!

1 Like