search

Zero-We / PMIL

Prototypical multiple instance learning for predicting lymph node metastasis of breast cancer from whole-slide pathological images
GNU General Public License v3.0
27 stars 4 forks source link

cluster stage question. #3

Open 16rq opened 9 months ago

16rq commented 9 months ago

Is there any suggetions to 'ConvergenceWarning: Affinity propagation did not converge, this model will not have any cluster centers.' in cluster.py?

MrPeterJin commented 9 months ago

I also encountered this problem. However I think it is depending on the original data, maybe the original WSI is too diverse to cluster (?)

Zero-We commented 9 months ago

Hi, regarding to the AP clustering, I have some suggestions.

  1. Considering the high time complexity of AP, it might be beneficial to reduce the number of patches through sampling before clustering.

  2. Patch features can significantly influence clustering results. If you just use the model pretrained on ImageNet to extract patch features, clustering may not be work. In this work, we actually pretrained the patch-level feature extractor using Max-Pooling MIL to obtain patch features, so it may be worth trying patch features obtained through self-supervised learning.