Histoboard - Pathology Foundation Model Benchmarks

Back to Benchmarks

Plismbench
pathology
robustness
H&E

Robustness benchmark from Owkin evaluating embedding consistency across scanners and staining variations using cosine similarity and top-10 retrieval accuracy.

15 models evaluated

4 tasks

Organs:

multi-organ

Detailed Results

Model	Average rank	Average metric	Cosine Similarity Embedding Consistency	Top-10 Cross-Scanner Scanner Robustness	Top-10 Cross-Staining Stain Robustness	Top-10 Cross-Scanner/Staining Combined Robustness
1H0-mini	1.75	0.541	0.800	0.864	0.318	0.183
2CONCH	3.00	0.498	0.846	0.752	0.241	0.155
3H-optimus-0	3.25	0.480	0.685	0.744	0.327	0.166
4Virchow2	3.50	0.464	0.777	0.609	0.306	0.163
5Midnight-12k	6.00	0.373	0.748	0.435	0.200	0.108
6Prov-GigaPath	7.75	0.333	0.570	0.592	0.118	0.054
7UNI2	7.75	0.332	0.591	0.501	0.190	0.046
8Kaiko ViT-B/8	8.00	0.325	0.764	0.346	0.147	0.045
9UNI	8.50	0.325	0.547	0.532	0.169	0.053
10GPFM	9.50	0.265	0.594	0.356	0.092	0.017
11PLIP	10.25	0.244	0.878	0.054	0.040	0.004
12Phikon	12.00	0.193	0.622	0.125	0.021	0.004
13Kaiko ViT-L/14	11.75	0.183	0.569	0.115	0.041	0.006
14Phikon-v2	13.50	0.164	0.557	0.064	0.030	0.003
15Hibou-L	13.50	0.147	0.490	0.061	0.030	0.008