Table 2 Stronger WSI and fusion baselines under the same protocol (Chordoma, N = 126)

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

Panel A: WSI-only baselines (Pearson r, mean ± sd over fivefolds)

WSI model

ERS–CAF

LR

H

Macro-r

ABMIL (standard MIL)

0.586 ± 0.061

0.563 ± 0.058

0.528 ± 0.057

0.559 ± 0.048

CLIP-guided MIL (ours)

0.637 ± 0.052

0.609 ± 0.061

0.585 ± 0.048

0.610 ± 0.046

CLAM

0.542 ± 0.055

0.519 ± 0.063

0.496 ± 0.060

0.519 ± 0.053

DSMIL

0.551 ± 0.058

0.527 ± 0.061

0.503 ± 0.058

0.527 ± 0.055

TransMIL

0.566 ± 0.056

0.545 ± 0.059

0.512 ± 0.056

0.541 ± 0.051

CTransPath + MIL aggregator

0.573 ± 0.053

0.550 ± 0.057

0.521 ± 0.055

0.548 ± 0.049

Panel B: Image-to-transcriptomics baseline (HE2RNA-style)

Model

ERS–CAF

LR

H

Macro-r

HE2RNA-style WSI → expr → score

0.498 ± 0.064

0.481 ± 0.068

0.460 ± 0.065

0.480 ± 0.061

Panel C: Fusion alternatives (Pearson r, mean ± sd over 5 folds)

Fusion model

ERS–CAF

LR

H

Macro-r

Late fusion (MRI + WSI avg)

0.601 ± 0.050

0.579 ± 0.055

0.551 ± 0.053

0.577 ± 0.049

Concat + MLP (no gating)

0.625 ± 0.048

0.598 ± 0.052

0.568 ± 0.050

0.597 ± 0.046

Transformer mid-level fusion

0.632 ± 0.046

0.605 ± 0.050

0.576 ± 0.048

0.604 ± 0.045

Fusion (no CORAL)

0.681 ± 0.043

0.652 ± 0.049

0.620 ± 0.044

0.651 ± 0.043

Fusion (ours: gated + CORAL)

0.712 ± 0.047

0.685 ± 0.041

0.659 ± 0.039

0.685 ± 0.041

  1. All methods use the same patient-level fivefold CV splits, tiling/QC, and metrics as Table 1.
  2. Bold values indicate the best performance under each evaluation metric.
Back to article page