-
Notifications
You must be signed in to change notification settings - Fork 10
View Embed Page
After uploading images and making patches, View Embed option gets activated. On View Embed page user would be able to Train DL model and Embed Patches. Begin labelling the objects.
- Train DL : Is unsupervised deep-learning model training of the patches, to group them based on the morphology, color and other properties of the objects. Note stain of the image plays a major role here, hence stain normalization is a very important.
- Embed Patches : Generates a UMAP in a 2D space for the patches, and user will see a 2 dimensional scatter plot being visible on the screen.
Once the plot is generated, user can quickly check the distribution of objects in the plot using Show Patches option. This helps decide which area to begin labelling at. User would lasso on the selected area and grid plot will be populated with all the patches and labelling options will be enabled.
User has options to filter the plot correctly to be able to pick the patches they want to see and label efficiently.
After labelling a small amount, user should re train the deeplearning model, which will consider the labels and do a semi supervised training, and a re-embed will have the clusters more enhanced and grouped together further helping the user labelling. A quick view of show patches show the distribution after iterative training and embedding.
PS also has a feature called Search Patches, this feature helps search similar patches based on the image provided by the user.
PS's Wiki is complete documentation that explains to user how to use this tool and the reasons behind. Here is the catalogue for PS's wiki page:
Home: