search

binli123 / dsmil-wsi

DSMIL: Dual-stream multiple instance learning networks for tumor detection in Whole Slide Image
MIT License
377 stars 88 forks source link

Generate features for another dataset #91

Closed ritamartinspereira closed 8 months ago

ritamartinspereira commented 9 months ago

Hello, I was wondering if you could describe how you generated the pre-computed features provided in this repository so that i can reproduce the same thing using a different dataset. Thank you

binli123 commented 8 months ago

Hello, I believe you can find instructions in the readme https://github.com/binli123/dsmil-wsi?tab=readme-ov-file#training-on-your-own-datasets and https://github.com/binli123/dsmil-wsi?tab=readme-ov-file#processing-raw-wsi-data

ericwang10 commented 4 months ago

Hi,

Sorry for reopening this thread, but I had a few questions about training my own dataset.

Step 2: Crop patches once we crop our WSI files into patches, do we need to manually separate them into training and testing datasets?

Step 3: Train an embedder Is the embedder only trained on the training dataset? Would we need to train another embedder for the test dataset or do we just use the original trained embedder?

If the tissue/classification problem is different, should I train my own embedder, or could I still use the one you provided? For example, using the Camelyon16 embedder from breast cancer lymph node tissue to identify epilepsy in brain tissue (different tissue and classification problem)

Lastly, for cross-validation, would we have to train a separate embedder for each cross-validation fold since the training dataset is different?

Sorry if these questions are basic, I am new to these MIL techniques.

Thank you for your time!