quality of the prototype

[hljWmh]

May I ask, has there been a go at evaluating the quality of the prototype? Like how pure the clustering is, etc?

[hljWmh]

I use the divided prototype pkl to evaluate the patches of a single slide, and the evaluation metrics are Davies-Bouldin Index and Silhouette Score, but I find that these metrics have a worrying effect. Is the quality of the prototype not good enough?

[andrewsong90]

Hi @hljWmh

Thank you for interest in the work. Assuming you are talking about morphological prototypes, can you provide more details (e.g., which patch encoder, which cohort). Few quick tips are 1) Use K-means (sklearn no FAISS), 2) use the strong patch encoder (not ResNet50 or CTranspath), and 3) try de-meaning the patches within each WSI to see if it helps.

Hope this helps,
Andrew

[hljWmh]

Thank you for your reply, the encoder I used is UNI, evaluated on KIRC dataset, clustering I did the construction of prototype for k-means and faiss respectively, then after getting the prototype I used the extracted patch features for the prototype quality evaluation with two metrics respectively, but the effect demonstrated that the two metrics are very ineffective? I am curious to know if it is a problem with my assessment or what? May I ask how you assessed the quality of the prototype at that time? What metrics were used?