Back to Benchmarks

CRoMa
pathology
robustness
H&E

Cross-confounder robustness margin measuring how much of a model's embedding neighbourhood is driven by biology rather than by the acquisition center. Builds on PathoROB's Robustness Index and scores 25 encoders on its three tile cohorts, reporting a median margin alongside a lower-tail severity.

25 models evaluated
3 tasks
Organs:
multi-organ
breast
esophagus

Detailed Results

Showing 3 of 3 tile cohorts. CRoMa is the median signed margin at the headline radius m=5, in (-1, 1) and neutral at 0. Each cell shows it above its lower-tail mean LTM₁₀. Models are ordered by the mean of the CRoMa rank and the tail rank (lower is better) — a reading order, not a score.

Model
Mean
rank
CRoMa
rank
Tail
rank
Average
CRoMa
Camelyon
CRoMa
TCGA-4x4
CRoMa
Tolkach ESCA
CRoMa
1MascaretTCGA
1.72.01.30.354
0.285tail -0.016
0.270tail -0.109
0.507tail 0.009
4.32.36.30.334
0.324tail -0.024
0.192tail -0.163
0.485tail -0.003
4.54.34.70.275
0.243tail -0.040
0.168tail -0.124
0.412tail -0.011
5.54.76.30.270
0.235tail -0.047
0.169tail -0.137
0.407tail -0.024
7.57.37.70.244
0.187tail -0.145
0.153tail -0.130
0.392tail -0.029
8.57.010.00.246
0.191tail -0.069
0.155tail -0.193
0.390tail -0.022
9.28.310.00.226
0.199tail -0.106
0.128tail -0.170
0.351tail -0.037
9.26.012.30.261
0.196tail -0.202
0.146tail -0.152
0.439tail -0.041
9.312.76.00.143
0.082tail -0.138
0.088tail -0.103
0.260tail -0.042
10PhaetTCGA
10.012.77.30.147
0.111tail -0.177
0.081tail -0.110
0.249tail -0.030
11H0-miniTCGA
12.09.314.70.222
0.167tail -0.160
0.121tail -0.188
0.380tail -0.074
12.210.314.00.206
0.155tail -0.181
0.092tail -0.180
0.371tail -0.055
12.24.719.70.363
0.108tail -0.349
0.396tail -0.210
0.584tail -0.083
12.315.39.30.110
0.045tail -0.154
0.054tail -0.111
0.232tail -0.080
13.315.311.30.114
0.045tail -0.209
0.075tail -0.123
0.221tail -0.063
16mSTARTCGA
14.316.712.00.090
0.022tail -0.177
0.059tail -0.124
0.187tail -0.089
17MUSKTCGA
14.515.014.00.129
0.042tail -0.220
0.051tail -0.140
0.293tail -0.066
18UNI
16.019.312.70.063
-0.034tail -0.217
0.047tail -0.120
0.175tail -0.080
17.319.315.30.064
0.009tail -0.190
0.051tail -0.151
0.132tail -0.160
20GPFMTCGA
18.319.017.70.058
-0.104tail -0.361
0.037tail -0.153
0.241tail -0.100
20.521.319.70.027
-0.089tail -0.363
0.036tail -0.177
0.134tail -0.167
22PhikonTCGA
21.721.022.30.005
-0.196tail -0.476
0.036tail -0.200
0.175tail -0.188
23Phikon-v2TCGA
22.523.721.3-0.019
-0.208tail -0.496
0.029tail -0.187
0.122tail -0.175
24Prost40MTCGA
24.024.024.0-0.067
-0.318tail -0.640
-0.010tail -0.250
0.127tail -0.236
24.223.325.0-0.097
-0.443tail -0.659
0.039tail -0.252
0.111tail -0.291

marks the Pareto frontier — the encoders no other pathology encoder beats on both the CRoMa rank and the tail rank at once. It is a set, not an order.

TCGA marks encoders whose disclosed pretraining overlaps TCGA, one of the three cohorts behind these ranks.

CRoMa also scores DINOv2-B as a natural-image calibration floor and PRISM/PRISM2/TITAN/MOOZY on a separate slide-level cohort. Neither takes part in the ranks above, and generic vision baselines are out of scope for Histoboard — see the full panel.