GSoC - Coding Period Week 1

June 14, 2021

This blog post showcases what I did last week and what I plan to do next.

Work Done This Week (June 7th to June 13th)

Last week I was able to convert the .tif files from the cell-tracking-challenge dataset into NumPy arrays. The next goal was to save these NumPy arrays as image files.
Here’s where I made a mistake, I saved the images as .jpeg files.
What I didn’t know is that JPEG is a lossy compression method, it changes pixel level values as long as it “looks good” to the human observer.
I used plotly to zoom into the saved .jpeg files, and noticed that the edges of the segmentation masks were noisy, then I plotted the images in 3D, with z axis as pixel value.

Turns out, JPEG is a lossy compression method - it is designed to save images using less bits even if it changes the pixel values as long as it “looks good” to a human observer.
I decided to save the images in .png format; PNG is a lossless compression method.
Each .tif file contains 3D data (as slices) from a single instant of time, I’ve named the corresponding .png files accordingly.
- features/F{time-point}_{slice}.png
- segmentation_maps/L{time-point}_{slice}.png
Link to code:
- Github
- Google Colab

I did not want to use torchvision.datasets.ImageFolder() for the dataset class.
The filenames (both features and labels) contain the time-point and slice number, which could be used to identify unique feature-label pairs.
I created a Pandas DataFrame that contains these unique id’s and saved it as a csv file, using which the custom dataloader loads valid feature-label pairs.

Gradio has been all over the ML community lately, I decided to give it a try.
The Gif below showcases a simple GUI (prototype) for the existing DevoLearn embryo segmentation model.

This still needs a lot more work. I’ll try to resume work on this after I’m done with the proposed work for next week.