Tutorial for the MicroUS Prostate Dataset

[TheHuntsman4]

Adds a simple segmentation tutorial for the Micro US Prostate dataset.
Link to the dataset: https://zenodo.org/records/10475293

The tutorial covers:

1. Downloading the dataset.
2. How to use the dataset with MONAI Dataloaders and Datasets.
3. Data transforms for training and validation
4. UNet model setup for segmentation
5. Training loop with validation
6. Visualization of results
7. Model evaluation

[Sharpz7]

Good overall:

• Lets try and get the results to be slightly more meaningful (e.g. when comparing the labels vs prediction have it working without all the artifacts)
• For the auto-downloading idea (i.e. handling the data downloading/processing entirely from the notebook) I feel it depends on the opinions of the MONAI maintainers + personally, I would make the dividing line whether it is using zenodo or something else that we trust to stay online.

I think it comes down to weighing stability vs smart

Open Questions:

• The data is huge, we need a way to down sample

[Sharpz7]

Since the "volume" is actually a 2D Sweep (i.e. a video) you can pass in the images to the model frame by frame.

So instead of a 1372, 962, 45 volume, pass in 1372, 962 slices as "images" to the model. This should stop you needing to downsize, speed up model training (less computation) and maybe even increase the batch size :)

I think we should downsize, I think that might be important for this dataset. Maybe consider moving duplicates from the videos too (https://files.mcaq.me/90697.png)

[gauthamk02]

It will be good to add an "open in collab" badge to the notebook.
https://openincolab.com/

I think the notebook can also named based on the task it does, something like 2D-US-segmentation-prostate.ipynb.

Just my suggestions. Totally fine to skip them if they’re not practical.