Perform TSNE on Segmentation Masks

[JarrodRShepherd]

Description

Perform TSNE on the watershed segmentation masks to determine if these separate better than the raw images. If they do, then we can go ahead and train a U-Net model on these watershed segmentation masks. If we are not satisfied with the results from TSNE, then we can train an unsupervised segmentation model and determine if these masks have better separation.

Files

• src\dataset\bps_datamodule.py from Homework 3
• src\dataset\bps_dataset.py from Homework 3
• src\model\unsupervised\pca_tsne.py from Homework 3
• src\vis_utils.py from Homework 3

Checklist

• [x] Port over the files from Homework 3
• [x] Adapt these files for our implementation
• [x] Run TSNE on our watershed segmentation masks
• [x] Analyze the results form TSNE
• [ ] Determine if we need to train an unsupervised model and repeat TSNE analysis

[JarrodRShepherd]

These are the outputs of TSNE on our watershed masks.

These results are not great, as TSNE was not able to easily separate the two classes. We then experimented with different thresholding techniques and watershed parameters. The results from these further experiments were better than the images shown here, but still did not produce good separable results.

Thinking about future steps, we have two options. The first is to iterate through many watershed parameters to determine which produces separable TSNE results. The second is to train an unsupervised model and then analyze the TSNE results of the masks produced by this model.