Table 3 External generalization and deployability

From: Decoding the ERS–CAF immunoregulatory axis via multimodal AI and its pan-cancer prognostic and therapeutic predictive value

Panel A: Pan-cancer zero-shot concordance (TCGA)

Cancer

ERS–CAF (r)

LR (r)

H (r)

PAAD

0.42 ± 0.03

0.39 ± 0.03

0.36 ± 0.04

STAD

0.38 ± 0.04

0.35 ± 0.04

0.33 ± 0.04

COADREAD

0.35 ± 0.04

0.31 ± 0.05

0.29 ± 0.05

Panel B: MRI-only distillation (Chordoma)

 

Macro-r

Retention

Speed/mem.

Teacher (multimodal)

0.68 ± 0.03

–

1 × /1 ×

Student (MRI-only)

0.64 ± 0.03

94%

4.6 × /2.9 ×

  1. Panel A: zero-shot pan-cancer concordance (Pearson r) between image-derived and transcriptomic surrogates. Panel B: MRI-only student vs. multimodal teacher on the chordoma cohort (macro-average across targets). In TCGA, concordance is computed using the frozen model with WSI input only (radiology withheld) due to heterogeneous radiology availability.
Back to article page