acceptodds
Under review as a conference paper at ICLR 2027

Ordinal Privileged Distillation for Histopathology-Only Survival Prediction

Abstract

Multimodal learning can improve cancer survival prediction, but some modalities may be unavailable at deployment. Existing approaches often address this limitation through molecular reconstruction or representation matching, which can require learning variation beyond the survival task. We instead use a teacher trained on the auxiliary modality alone and take its survival outputs as the sole source of supervision from that modality. Transferring these outputs raises a further challenge. Existing survival distillation objectives typically encourage numerical agreement, although teachers and students observing different modalities need not share the same optimal predictions. We prove that matching the teacher’s event probabilities at each time interval can shift the student’s optimal predictions. We therefore introduce Ordinal Privileged Distillation (OPD), which transfers pairwise risk ordering derived from the teacher’s survival curves without matching its absolute score scale or pairwise score gaps. A pairwise logistic loss supplies ordinal supervision alongside a survival loss that accounts for right censoring, leaving the student’s inference architecture unchanged. We evaluate OPD using transcriptome data as the training-only modality and whole-slide images (WSIs) as the student’s input. Extensive experiments on eight cancer benchmark datasets and four WSI backbones demonstrated our superiority over other methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Computational Pathology, Whole Slide Image, Cancer Survival Prediction

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_dmXvOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies the problem of unvailable modalities in the deployment of survival prediction methods. Other approaches address this problem by reconstruction or presentation matching. This may require learning variation beyond the survival task. The authors used a teacher trained on the auxiliary modality alone and take its survival outputs as the sole source of supervision from that modality and figured that existing survival distillation objectives typically encourage numerical agreement, although teachers and students observing different modalities need not share the same optimal predictions. They then prove that matching the teacher’s event probabilities at each time interval can shift the student’s optimal predictions. The authors then propose Ordinal Privileged Distillation (OPD), this method can transfer pair-wise rish ordering derived from the teacher's survival curves without matching its absolute score scale or pairwise score gaps. A pairwise logistic loss supplies ordinal supervision alongside a survival loss that accounts for right censoring, leaving the student’s inference architecture unchanged. Experimental results shows the method is effective.

Critical strengths
  1. The idea is simple but effective. Th student learns pairwise risk ordering from single modal RNA teacher by pairwise loogistic loss. It does not introduce any incremental changes in network structure or training framework, instead, it is a purely loss-driven distillation paradigm that operates strictly at the output risk level.

  2. The authors insightfully observe that 'the teacher and student condition on different inputs... so their optimal predictions can differ'. Proposition 2 proves that output matching requires an unmet compatibility condition which naturally motivates ordinal supervision that is 'invariant to strictly increasing transformations' without forcing the student to match 'numerical risk gaps'.

Critical weaknesses
  1. The pairwise loss scales quadratically within each batch, relying on an effective batch size. The paper currently lacks ablation experiments on sensitivity across varying effective batch sizes.
Questions for authors
  1. The pairwise loss scales as O(B2)\mathcal{O}(B^2) and is tested with an effective batch size of B=64B=64. In memory-constrained scenarios where the batch size is much smaller (e.g., B=8B=8 or 1616), does the reduction in pair diversity degrade the batch standardization and ordinal supervision? Have the authors evaluated the model's sensitivity to smaller batch sizes?

  2. Since the RNA teacher is trained on right-censored data, how reliable are its continuous risk scores for pairs with very short follow-up times and early censoring? Would adding a confidence threshold or margin filter—to supervise only high-certainty pairs—improve stability?

3 comments

Thank you for the positive evaluation and the practical questions.


Q1. The pairwise loss scales as O(B2)O(B^2) and is tested with an effective batch size of 64. In memory-constrained scenarios where the batch size is much smaller, does the reduction in pair diversity degrade the batch standardization and ordinal supervision? Have the authors evaluated the model's sensitivity to smaller batch sizes?

We retrained all four backbones, with and without OPD, at actual batch sizes of 8, 16, and 32, and compared them with the main batch-64 results:

Batch sizeABMIL (NLL → OPD)TransMIL (NLL → OPD)MMP-Histo (NLL → OPD)PANTHER (NLL → OPD)Mean gain (points)
80.6173 → 0.66580.6350 → 0.66490.6375 → 0.69510.6581 → 0.6861+4.10
160.6204 → 0.66700.6439 → 0.66980.6391 → 0.69600.6560 → 0.6883+4.04
320.6179 → 0.66780.6433 → 0.66900.6401 → 0.69910.6573 → 0.6865+4.10
64 (main)0.6196 → 0.66550.6435 → 0.66990.6416 → 0.69940.6570 → 0.6866+3.99

For each backbone, the result with OPD varies by at most 0.5 points across batch sizes, and the gain over NLL-only training at the same batch size stays at about 4 points. The ordinal signal and the batch standardization remain effective.

Full per-cohort results for the batch-size study. BB is the actual batch size; B=64B = 64 is the main setting.

ModelBBBRCACRCGBMLGGHNSCKIRCLUADSKCMUCECMean
ABMIL80.645±0.0500.550±0.0830.763±0.0380.573±0.0800.657±0.0600.550±0.0330.557±0.0530.643±0.2040.617
ABMIL160.644±0.0500.559±0.0840.760±0.0420.577±0.0840.658±0.0610.558±0.0370.558±0.0510.649±0.1860.620
ABMIL320.641±0.0570.555±0.0880.761±0.0420.571±0.0770.656±0.0670.553±0.0360.557±0.0460.649±0.1970.618
ABMIL640.644±0.0580.557±0.1030.767±0.0440.577±0.0580.654±0.0730.556±0.0400.555±0.0430.647±0.1910.620
ABMIL + OPD80.660±0.0550.645±0.0680.787±0.0300.631±0.0650.676±0.0680.596±0.0310.620±0.0470.711±0.1450.666
ABMIL + OPD160.665±0.0500.642±0.0690.783±0.0320.633±0.0700.678±0.0700.593±0.0310.623±0.0460.719±0.1630.667
ABMIL + OPD320.668±0.0500.642±0.0650.783±0.0310.637±0.0760.675±0.0740.597±0.0280.623±0.0450.717±0.1700.668
ABMIL + OPD640.665±0.0560.640±0.0930.780±0.0340.636±0.0790.676±0.0750.591±0.0290.623±0.0430.713±0.1760.666
TransMIL80.641±0.0700.636±0.0540.765±0.0420.613±0.0430.666±0.0880.573±0.0420.582±0.0360.604±0.1780.635
TransMIL160.677±0.0740.632±0.0490.766±0.0430.615±0.0450.665±0.0860.574±0.0390.576±0.0420.646±0.1780.644
TransMIL320.672±0.0490.638±0.0420.762±0.0460.612±0.0310.663±0.0920.577±0.0240.584±0.0470.638±0.1870.643
TransMIL640.676±0.0600.639±0.0350.764±0.0510.612±0.0390.661±0.0910.579±0.0340.583±0.0520.634±0.1870.644
TransMIL + OPD80.668±0.0650.656±0.0170.775±0.0430.623±0.0290.683±0.0720.606±0.0310.606±0.0390.702±0.1190.665
TransMIL + OPD160.689±0.0740.663±0.0140.777±0.0360.623±0.0300.681±0.0750.606±0.0360.604±0.0460.715±0.1220.670
TransMIL + OPD320.687±0.0470.666±0.0300.774±0.0360.626±0.0280.682±0.0770.605±0.0360.602±0.0430.710±0.1380.669
TransMIL + OPD640.683±0.0480.664±0.0240.777±0.0420.622±0.0410.682±0.0760.607±0.0390.607±0.0430.717±0.1490.670
MMP-Histo80.625±0.1100.671±0.1130.733±0.0400.592±0.0550.627±0.0780.589±0.0540.599±0.0630.664±0.1830.638
MMP-Histo160.626±0.0990.673±0.0890.733±0.0590.611±0.0600.626±0.0800.583±0.0620.601±0.0620.660±0.1940.639
MMP-Histo320.629±0.1010.671±0.1000.731±0.0980.610±0.0440.623±0.0780.589±0.0540.604±0.0730.664±0.1600.640
MMP-Histo640.630±0.0600.674±0.1040.732±0.1030.609±0.0550.626±0.0570.586±0.0550.610±0.0780.666±0.1580.642
MMP-Histo + OPD80.683±0.0590.693±0.0620.777±0.0360.642±0.0810.711±0.0600.620±0.0600.641±0.0650.794±0.1320.695
MMP-Histo + OPD160.680±0.0630.692±0.0770.780±0.0440.641±0.0690.710±0.0610.628±0.0620.651±0.0750.786±0.1510.696
MMP-Histo + OPD320.682±0.0730.691±0.0870.789±0.0410.647±0.0730.713±0.0620.631±0.0580.645±0.0800.795±0.1340.699
MMP-Histo + OPD640.683±0.0690.699±0.0810.783±0.0330.640±0.0730.713±0.0620.630±0.0570.651±0.0790.796±0.1090.699
PANTHER80.685±0.0950.689±0.1030.763±0.0380.588±0.0620.686±0.0640.613±0.0450.617±0.0490.624±0.1690.658
PANTHER160.681±0.1000.684±0.1130.768±0.0360.581±0.0690.686±0.0650.618±0.0500.610±0.0570.620±0.1910.656
PANTHER320.689±0.0980.676±0.1230.768±0.0340.584±0.0760.684±0.0690.616±0.0510.616±0.0620.625±0.1510.657
PANTHER640.683±0.1100.680±0.1040.764±0.0310.586±0.0750.683±0.0690.619±0.0480.618±0.0640.623±0.1090.657
PANTHER + OPD80.680±0.0670.693±0.1060.781±0.0280.633±0.0460.717±0.0530.627±0.0560.640±0.0290.718±0.1590.686
PANTHER + OPD160.686±0.0800.696±0.0920.784±0.0340.636±0.0530.716±0.0540.628±0.0560.642±0.0240.718±0.1690.688
PANTHER + OPD320.683±0.0700.692±0.1000.788±0.0330.634±0.0520.714±0.0550.622±0.0610.642±0.0170.717±0.1950.687
PANTHER + OPD640.688±0.0810.690±0.1040.782±0.0310.631±0.0480.710±0.0540.625±0.0540.648±0.0260.719±0.2130.687

Q2. Since the RNA teacher is trained on right-censored data, how reliable are its continuous risk scores for pairs with very short follow-up times and early censoring? Would adding a confidence threshold or margin filter—to supervise only high-certainty pairs—improve stability?

A censoring-aware training likelihood does not guarantee that every teacher comparison is reliable. In particular, an early-censored outcome may not establish a patient's order relative to another patient, whereas the teacher can still predict an order from their RNA profiles. We tested the strictest high-certainty filter that the outcomes allow: keeping only Harrell-comparable pairs, in which the patient with the earlier observed time must have an observed event. This excludes censored–censored pairs and pairs whose earlier observation is censored, while keeping the RNA teacher as the source of the comparison labels:

Model (pairs used)BRCACRCGBMLGGHNSCKIRCLUADSKCMUCECMeanMean std
ABMIL (Harrell pairs)0.665±0.0560.646±0.0820.776±0.0420.644±0.0740.675±0.0780.591±0.0310.621±0.0450.672±0.1920.6610.075
ABMIL (all pairs)0.665±0.0560.640±0.0930.780±0.0340.636±0.0790.676±0.0750.591±0.0290.623±0.0430.713±0.1760.6660.073
TransMIL (Harrell pairs)0.682±0.0510.662±0.0270.775±0.0490.634±0.0460.679±0.0780.606±0.0310.602±0.0350.695±0.1400.6670.057
TransMIL (all pairs)0.683±0.0480.664±0.0240.777±0.0420.622±0.0410.682±0.0760.607±0.0390.607±0.0430.717±0.1490.6700.058
MMP-Histo (Harrell pairs)0.682±0.0710.685±0.0960.783±0.0600.639±0.0740.700±0.0730.633±0.0500.647±0.0780.706±0.1620.6840.083
MMP-Histo (all pairs)0.683±0.0690.699±0.0810.783±0.0330.640±0.0730.713±0.0620.630±0.0570.651±0.0790.796±0.1090.6990.070
PANTHER (Harrell pairs)0.694±0.0940.689±0.0900.781±0.0340.629±0.0450.707±0.0610.617±0.0520.637±0.0290.689±0.2130.6800.077
PANTHER (all pairs)0.688±0.0810.690±0.1040.782±0.0310.631±0.0480.710±0.0540.625±0.0540.648±0.0260.719±0.2130.6870.076

Restricting OPD to Harrell-comparable pairs does not improve accuracy or stability: performance changes little on seven cohorts and drops noticeably on UCEC (0.672 vs. 0.713 with ABMIL; 0.706 vs. 0.796 with MMP-Histo). The reason is fundamental to privileged distillation: even when a patient is right-censored early in follow-up, their transcriptomic profile GiG_i is fully observed. Because the RNA teacher maps GiG_i to a full survival curve and is trained across the cohort with a censoring-aware NLL, its continuous risk score rig=−∑k=1ItSig,(k)r_i^g = -\sum_{k=1}^{I_t} S_i^{g,(k)} provides an informative molecular risk ordering for censored pairs whose raw follow-up times alone are uninformative, as in UCEC (n=149n = 149).

We hope this addresses your concerns, and we are glad to answer any further questions.

Reviewer_JjzXOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PANTHER, a cross-modal knowledge distillation framework designed to address the limitations of using a single visual modality for cancer survival prediction in computational pathology. Its main contribution lies in the innovative design of survival analysis-specific distillation objectives and loss functions, which enable the transfer of molecular knowledge, including RNA-seq transcriptomic features, from a teacher model to an image-based student model, thereby effectively bridging the semantic gap between morphological and molecular features. Regarding validation, the paper employs standardized WSI and RNA preprocessing, rigorous cross-validation, and detailed ablation studies. The experimental results demonstrate that PANTHER significantly outperforms baseline methods across multiple evaluation metrics, highlighting its strong robustness and generalization ability in complex clinical survival prediction tasks.

Critical strengths

The most critical strength of this paper lies in the consistent and highly significant survival risk stratification demonstrated across multiple TCGA cohorts (e.g., COAD/READ and the cohort with n = 525), with most log-rank p-values below 0.001. Furthermore, the ablation studies demonstrate that the OPD mechanism is highly robust to variations in the distillation weight λ, with C-index fluctuations of less than 0.01 across λ ∈ {0.3, 1, 3, 10}. This strength is particularly important to my overall assessment because it provides strong clinical evidence that cross-modal ordinal distillation from a neural RNA teacher can effectively address the limitations of WSI-only models in survival prediction. Moreover, from an algorithmic perspective, these results demonstrate that the distillation signal is stable and reproducible, rather than relying on fragile hyperparameter tuning.

Critical weaknesses

The interpretation of the attention mechanism relies heavily on established clinical and pathological prior knowledge, such as lymphocytic infiltration in SKCM and solid growth patterns in LUAD. The paper does not demonstrate that OPD genuinely captures deeper cross-modal survival mechanisms from the RNA teacher beyond conventional pathological knowledge. In addition, although the paper claims consistent improvements across multiple cohorts, the experimental results suggest that its relative advantages are somewhat marginal. For example, several baseline models in the CRC cohort also achieve log-rank p-values below 0.001. Furthermore, all validation experiments are limited to TCGA data, leaving the generalizability of the proposed method beyond TCGA insufficiently established.

Questions for authors
  1. Could the authors clarify the specific mathematical assumptions in Appendix A under which OPD's pairwise risk ranking is theoretically guaranteed to outperform direct score matching methods, such as RankDistil?
  2. Could the authors clarify whether the performance improvements achieved by OPD in cohorts where baseline models also perform strongly have been evaluated using rigorous statistical significance tests (e.g., paired t-tests) to rule out the possibility that the observed gains are attributable to random variation?
2 comments

Thank you for the review. First, a clarification on naming: our method is Ordinal Privileged Distillation (OPD). PANTHER (Song et al., 2024a) is one of the four WSI backbones to which we apply OPD (Table 1).


Weakness. The attention analysis relies on known pathology.

We do not claim mechanisms beyond established pathology. A WSI-only student can use only morphology, so supervision derived from RNA can act only by changing which morphology the student relies on. Figure 2 illustrates how the WSI regions emphasized by the student change after RNA-derived supervision is added. That these shifts agree with known prognostic patterns (tumor-infiltrating lymphocytes in melanoma, solid growth in lung adenocarcinoma) is the outcome one would hope for clinically.


Weakness. Generalizability beyond TCGA.

External validation on CPTAC cohorts. We added results on CPTAC. Among our eight cohorts, only four have enough CPTAC samples with paired WSI–RNA data. The results are shown below (bold: OPD outperforms all five existing methods).

ModelLUADHNSCUCECKIRC
Existing methods (RNA during training)
G-HANet0.631±0.0590.629±0.1080.616±0.1730.664±0.076
ProSurv0.556±0.0850.635±0.1360.670±0.0810.628±0.183
LD-CVAE0.630±0.0550.625±0.0350.662±0.0900.659±0.101
DisPro0.562±0.0700.619±0.1090.635±0.1130.670±0.142
SlotSPE0.617±0.0830.664±0.1200.631±0.1610.638±0.084
WSI backbones: NLL only vs. + OPD
ABMIL0.568±0.1070.540±0.1990.622±0.1870.619±0.095
ABMIL + OPD0.632±0.1300.610±0.1670.642±0.1620.642±0.095
TransMIL0.584±0.0780.619±0.1020.615±0.1540.647±0.147
TransMIL + OPD0.609±0.0730.662±0.0740.640±0.1500.678±0.168
PANTHER0.531±0.0800.643±0.1610.542±0.0870.645±0.106
PANTHER + OPD0.571±0.1090.661±0.1120.631±0.1060.692±0.074
MMP-Histo0.595±0.1540.649±0.1950.606±0.1080.586±0.149
MMP-Histo + OPD0.598±0.0690.668±0.0710.723±0.0830.685±0.079

OPD improves every backbone on every cohort (16/16 settings), by 4.6 C-index points on average, close to the 4.0 points on TCGA. In each cohort, the best OPD student is above the best existing method that uses RNA during training: narrowly on LUAD and HNSC (+0.1 and +0.4 points) and clearly on UCEC and KIRC (+5.3 and +2.2 points).


Q1. Could the authors clarify the specific mathematical assumptions in Appendix A under which OPD's pairwise risk ranking is theoretically guaranteed to outperform direct score matching methods, such as RankDistil?

We do not claim a theoretical guarantee that OPD outperforms direct score-matching methods. Appendix A.3 states that the analysis motivates relaxing numerical agreement "without proving that teacher comparisons are always informative or that ranking must outperform KL". What Appendix A does establish, and under which assumptions:

  • Proposition 1 (A.1). For the discrete-time endpoint of Section 3.1, the ideal RNA survival curve determines the conditional event distribution given RNA. No further assumption is needed. It says nothing about a learned teacher or about the scalar risk score.
  • Proposition 2 (A.3). Assume censoring that is independent of the event time given the WSI, a fixed teacher, and the untempered single-event interval-wise KL with the same at-risk weights as the survival loss. Then the population-optimal student hazard is pulled toward the average teacher hazard (Eq. 8). Hazard matching leaves the student's optimum unchanged only under the compatibility condition stated there, which paired data do not guarantee.
  • OPD. Its supervision is invariant to any strictly increasing transformation of the teacher's scores, so it does not need that condition. In exchange, it inherits the teacher's ordering errors: a chance-level teacher (our linear RNA teacher, held-out C-index 0.52) lowers the student by 6.4 points. RankDistil (Reddi et al., 2021) is itself a ranking-distillation method from the learning-to-rank literature, and it cannot handle right-censored survival data. We cite it as related work that shares a similar ranking-distillation idea.

Q2 / Weakness. Could the authors clarify whether the performance improvements achieved by OPD in cohorts where baseline models also perform strongly have been evaluated using rigorous statistical significance tests (e.g., paired t-tests) to rule out the possibility that the observed gains are attributable to random variation?

Paired statistical significance tests. We evaluated OPD with two-sided paired tt-tests and non-parametric Wilcoxon signed-rank tests across cohorts and backbones.

(a) OPD vs. no distillation (TCGA, 32 backbone–cohort settings). OPD improves the C-index in all 32 settings (W/T/L = 32/0/0), with a mean gain of ΔC×100=+3.99±1.10\Delta C \times 100 = +3.99 \pm 1.10 (95% CI); paired t=7.41t = 7.41, p=2.41×10−8p = 2.41 \times 10^{-8} (Wilcoxon p=4.66×10−10p = 4.66 \times 10^{-10}). Within each individual backbone (n=8n = 8 cohorts), the gain is also statistically significant: ABMIL (+4.59+4.59, p=0.0017p = 0.0017), TransMIL (+2.64+2.64, p=0.0175p = 0.0175), MMP-Histo (+5.78+5.78, p=0.0022p = 0.0022), and PANTHER (+2.96+2.96, p=0.0271p = 0.0271).

(b) OPD vs. existing paired WSI–RNA methods (TCGA, n=8n = 8 cohorts). We compare our top two OPD students against all five state-of-the-art missing-modality methods (G-HANet, ProSurv, LD-CVAE, DisPro, SlotSPE):

OPD studentStatisticvs. G-HANetvs. ProSurvvs. LD-CVAEvs. DisProvs. SlotSPE
MMP-Histo + OPDΔC×100\Delta C \times 100+4.78+7.74+3.55+9.60+5.31
W/T/L8/0/08/0/07/0/18/0/08/0/0
Paired tt-test pp / Wilcoxon pp0.019 / 0.0080.004 / 0.0080.062 / 0.0160.002 / 0.0080.050 / 0.008
PANTHER + OPDΔC×100\Delta C \times 100+3.50+6.46+2.27+8.32+4.04
W/T/L8/0/08/0/07/0/18/0/08/0/0
Paired tt-test pp / Wilcoxon pp0.006 / 0.0080.0005 / 0.0080.027 / 0.0160.0001 / 0.0080.024 / 0.008

Across the 40 cohort-level comparisons against the five existing methods, MMP-Histo + OPD and PANTHER + OPD each win 39 (p≤0.027p \le 0.027 against all five methods for PANTHER + OPD with the paired tt-test, and p≤0.016p \le 0.016 against all five methods for both students with the Wilcoxon test).

Strong cohorts, and why the Kaplan–Meier p<0.001p < 0.001 saturates. In the Kaplan–Meier plots (Figures 3–10), pp-values are truncated at p<0.001p < 0.001 whenever the log-rank statistic χ12\chi^2_1 exceeds 10.83, which masks substantial differences in fine-grained patient ranking (C-index):

  • On CRC (Figure 4), DisPro (p=0.015p = 0.015), G-HANet (p=0.025p = 0.025), ProSurv (p=0.005p = 0.005), and LD-CVAE (p=0.035p = 0.035) do not reach p<0.001p < 0.001, and the ABMIL baseline fails to stratify high- and low-risk patients (p=0.155p = 0.155, C-index 0.557). Adding OPD improves ABMIL by 8.3 C-index points (to 0.640, reaching p<0.001p < 0.001), while MMP-Histo + OPD (0.699) and PANTHER + OPD (0.690) both surpass the strongest existing method, SlotSPE (0.688).
  • Even on cohorts where all baselines already perform strongly (GBMLGG and KIRC), OPD consistently improves all four backbones (+1.3 to +5.1 points on GBMLGG; +2.1 to +8.7 points on KIRC) and outperforms every baseline.

We hope this addresses your concerns, and we are glad to answer any further questions.

Reviewer_fTznOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper targets WSI-only survival prediction when RNA is available only during training. A frozen RNA-only survival teacher turns its predicted curves into risk scores, and the WSI student gets a pairwise logistic loss that follows the teacher's risk ordering, added to the usual discrete-time survival loss. Two propositions argue that the survival curve is a sufficient target and that hazard matching shifts the student's optimum. Experiments cover eight TCGA cohorts and four MIL backbones.

Critical strengths
  1. The method is simple, cheap and architecture-agnostic. It only changes the training loss, needs no reconstruction module, and leaves inference unchanged. It improves all four backbones on all eight cohorts, and MMP-Histo with OPD beats the paired-modality baselines on six of the eight cohorts.

  2. The ablations are well chosen. Comparing pathway targets, teacher embeddings, several distillation objectives and several teachers, including a permuted-RNA control that keeps the score distribution but breaks patient correspondence, gives a fairly complete picture of where the gain comes from. Proposition 2 is a clear, easy-to-check argument for why hazard matching across modalities can bias the student.

Critical weaknesses
  1. Part of the gain may come from adding a ranking loss aligned with the metric. C-index is pairwise, and OPD trains the ordering of the evaluated risk score. The RNA teacher is trained on the same patients, and the paper does not say whether its training-set scores are in-sample or out-of-fold, so they may largely echo observed outcomes. The self-teacher result helps, but a teacher-free pairwise loss on observed pairs is the key missing control.

  2. The evidence for the ordinal part is thin. Risk MSE with the same teacher already gives 3.31 of the 3.99 C-index points, so ranking adds only 0.68 on average. Fold standard deviations are often far larger than the gains, for example a gain of 0.007 on BRCA with TransMIL against a deviation of 0.048, and UCEC deviations range from 0.11 to 0.21. Paired tests across folds and the RNA teacher's own C-index would help.

  3. The theory supports less than the framing suggests. Proposition 1 only states that the full survival curve determines the event distribution, and Proposition 2 shows hazard matching pulls the student toward the teacher, but OPD also changes the student's optimum and has no comparable guarantee, as the appendix admits. In practice a weak teacher is harmful, since the linear RNA teacher drops mean C-index by 6.4 points, about as much as permuted outputs.

Questions for authors

see weaknesses.

4 comments

We thank you for the insightful and rigorous review. We address all three concerns below.


W1. Part of the gain may come from adding a ranking loss aligned with the metric. C-index is pairwise, and OPD trains the ordering of the evaluated risk score. The RNA teacher is trained on the same patients, and the paper does not say whether its training-set scores are in-sample or out-of-fold, so they may largely echo observed outcomes. The self-teacher result helps, but a teacher-free pairwise loss on observed pairs is the key missing control.

Teacher scores. The scores are in-sample. In each fold, the RNA teacher is trained only on that fold's training patients, and the scores it provides for these patients are in-sample. Held-out patients never enter teacher or student training.

Teacher-free control. We added the key missing control, a teacher-free pairwise loss on observed pairs:

Loss added to NLLPair labelsABMILTransMILMMP-HistoPANTHERΔC×100\Delta C \times 100W/T/L
None (NLL only)–0.6200.6440.6420.657––
Outcome rankingObserved outcomes0.6380.6450.6690.663+1.3524/0/8
OPDRNA teacher0.6660.6700.6990.687+3.9932/0/0

ΔC\Delta C and wins/ties/losses (W/T/L) are computed against NLL-only training over the 32 backbone–cohort settings, as in Table 2.

The results show that a teacher-free ranking loss recovers about one third of the gain, and the rest requires the RNA teacher's ordering. The improvement is therefore tied to the patient-specific ordering from the auxiliary modality, not to the addition of a metric-aligned ranking term.

Full per-cohort results:

ModelBRCACRCGBMLGGHNSCKIRCLUADSKCMUCECMean
ABMIL (NLL only)0.644±0.0580.557±0.1030.767±0.0440.577±0.0580.654±0.0730.556±0.0400.555±0.0430.647±0.1910.620
ABMIL + outcome ranking0.640±0.0540.607±0.0520.771±0.0430.609±0.0710.679±0.0810.562±0.0400.587±0.0340.653±0.1460.638
ABMIL + OPD0.665±0.0560.640±0.0930.780±0.0340.636±0.0790.676±0.0750.591±0.0290.623±0.0430.713±0.1760.666
TransMIL (NLL only)0.676±0.0600.639±0.0350.764±0.0510.612±0.0390.661±0.0910.579±0.0340.583±0.0520.634±0.1870.644
TransMIL + outcome ranking0.679±0.0550.636±0.0440.768±0.0500.595±0.0360.665±0.0840.593±0.0440.581±0.0480.645±0.1520.645
TransMIL + OPD0.683±0.0480.664±0.0240.777±0.0420.622±0.0410.682±0.0760.607±0.0390.607±0.0430.717±0.1490.670
MMP-Histo (NLL only)0.630±0.0600.674±0.1040.732±0.1030.609±0.0550.626±0.0570.586±0.0550.610±0.0780.666±0.1580.642
MMP-Histo + outcome ranking0.678±0.0860.706±0.0930.768±0.0690.578±0.0600.697±0.0680.642±0.0490.616±0.0770.669±0.1530.669
MMP-Histo + OPD0.683±0.0690.699±0.0810.783±0.0330.640±0.0730.713±0.0620.630±0.0570.651±0.0790.796±0.1090.699
PANTHER (NLL only)0.683±0.1100.680±0.1040.764±0.0310.586±0.0750.683±0.0690.619±0.0480.618±0.0640.623±0.1090.657
PANTHER + outcome ranking0.689±0.1050.662±0.0870.775±0.0400.576±0.0620.715±0.0550.617±0.0590.624±0.0500.644±0.1700.663
PANTHER + OPD0.688±0.0810.690±0.1040.782±0.0310.631±0.0480.710±0.0540.625±0.0540.648±0.0260.719±0.2130.687

W2. The evidence for the ordinal part is thin. Risk MSE with the same teacher already gives 3.31 of the 3.99 C-index points, so ranking adds only 0.68 on average. Fold standard deviations are often far larger than the gains, for example a gain of 0.007 on BRCA with TransMIL against a deviation of 0.048, and UCEC deviations range from 0.11 to 0.21. Paired tests across folds and the RNA teacher's own C-index would help.

Why the gain over Risk MSE is small. Our Risk MSE standardizes the student and teacher scores within the batch. For z-scores,

1B∑i(rˉip−rˉig)2=2(1−ρP),\frac{1}{B}\sum_i \left(\bar r_i^p - \bar r_i^g\right)^2 = 2(1-\rho_P),

where ρP\rho_P is the within-batch Pearson correlation between the student and teacher risks. Risk MSE is thus itself a scale- and shift-free relational objective: it asks the student to be linearly correlated with the teacher, which already transfers most of the teacher's ordering. Risk MSE and OPD are therefore two instantiations of the principle in Section 3.3: transfer the teacher's risk relationally instead of matching its values. They sit at two levels of invariance:

  • Standardized Risk MSE (affine invariance): invariant to positive affine transformations arg+ba r^g + b (a>0a > 0) of the teacher's scores, but it still forces the student to match relative score gaps.
  • OPD (monotonic invariance): invariant to any strictly increasing transformation of the teacher's scores. It transfers only the comparison signs qijq_{ij} and is a surrogate for Kendall-τ\tau agreement with the teacher. Relaxing affine agreement to monotonic agreement consistently yields the extra +0.68+0.68 gain (30/2/0 W/T/L).

Why fold SDs are large, and what paired tests show. In 5-fold site-stratified cross-validation, each validation fold consists of held-out clinical institutions with different patient demographics, censoring rates, and sample sizes. For example, UCEC has n=149n = 149 patients in total, with ∼30\sim 30 patients and <10< 10 events per fold, which causes a high inter-fold SD for all methods. Because all methods are evaluated on identical folds, the between-fold variance cancels out in paired comparisons. Paired differences across all backbone–cohort settings confirm that both the overall OPD gain and the specific ordinal gain over Risk MSE are highly statistically significant:

ComparisonΔC×100\Delta C \times 100 (mean ± 95% CI)W/T/LPaired tt-testWilcoxon pp
OPD vs. no distillation+3.99 ± 1.1032/0/0t=7.41t = 7.41, p=2.41×10−8p = 2.41 \times 10^{-8}4.66×10−104.66 \times 10^{-10}
OPD vs. DiSKD+3.17 ± 1.1828/1/3t=5.48t = 5.48, p=5.37×10−6p = 5.37 \times 10^{-6}6.85×10−66.85 \times 10^{-6}
OPD vs. Risk MSE+0.68 ± 0.2230/2/0t=6.30t = 6.30, p=5.25×10−7p = 5.25 \times 10^{-7}1.63×10−61.63 \times 10^{-6}
OPD vs. Risk MSE (ABMIL only)+0.65 ± 0.547/1/0t=2.86t = 2.86, p=0.0245p = 0.02450.0178
OPD vs. Risk MSE (TransMIL only)+0.36 ± 0.287/1/0t=3.04t = 3.04, p=0.0190p = 0.01900.0178
OPD vs. Risk MSE (MMP-Histo only)+1.03 ± 0.608/0/0t=4.02t = 4.02, p=0.0050p = 0.00500.0078
OPD vs. Risk MSE (PANTHER only)+0.67 ± 0.498/0/0t=3.26t = 3.26, p=0.0138p = 0.01380.0078

C-index of the teachers. The C-index of the different teachers is reported below.

TeacherBRCACRCGBMLGGHNSCKIRCLUADSKCMUCECMeanStudent ΔC×100\Delta C \times 100
RNA-only (used by OPD)0.654±0.0340.617±0.0540.837±0.0520.587±0.0470.671±0.0240.641±0.0220.638±0.0290.600±0.1270.656+3.99
Multimodal MMP0.661±0.0470.677±0.1080.830±0.0470.602±0.0480.684±0.0250.657±0.0250.633±0.0420.592±0.0570.667+2.51
Linear RNA0.445±0.0820.487±0.0960.756±0.0370.471±0.1040.491±0.0480.511±0.0380.440±0.0500.548±0.2320.519−6.38

Based on these results, we make three observations:

  1. A weak teacher is harmful, while an accurate one helps. The linear teacher is below 0.5 in five cohorts, and distilling from it lowers the student there. In the one cohort where it is accurate (GBMLGG, 0.756), it improves all four backbones (Table 5).
  2. Teacher accuracy alone does not determine the gain. The multimodal teacher is more accurate than the RNA-only teacher (0.667 vs. 0.656) but yields a smaller student gain (+2.51 vs. +3.99). The gain grows as the teacher's input overlaps less with the student's (WSI only, then WSI+RNA, then RNA only), in line with our motivation for a teacher that sees only the auxiliary modality.
  3. Students are not bounded by their teacher. The four OPD students reach 0.666–0.699, against 0.656 for the teacher. The students combine the teacher's ordering with WSI evidence and outcome supervision instead of copying the teacher.

W3. The theory supports less than the framing suggests. Proposition 1 only states that the full survival curve determines the event distribution, and Proposition 2 shows hazard matching pulls the student toward the teacher, but OPD also changes the student's optimum and has no comparable guarantee, as the appendix admits. In practice a weak teacher is harmful, since the linear RNA teacher drops mean C-index by 6.4 points, about as much as permuted outputs.

Our theory and analysis motivate the choice of target rather than establish it. In detail:

  • Proposition 1 is a sufficiency property of the ideal RNA survival curve. It motivates the choice of the survival output as the supervision target (with no need for molecular reconstruction or representation matching). It gives no guarantee for a learned teacher, for the scalar risk score, or for its pairwise ordering (Appendix A.1).
  • Proposition 2 characterizes when hazard matching preserves the student's optimum (the compatibility condition stated with Eq. 8). It describes a limitation of numerical matching across modalities, not a guarantee for OPD.
  • OPD has no comparable optimality guarantee (Appendix A.3). What it has is an invariance: its supervision is unchanged by any strictly increasing transformation of the teacher's scores, so it does not need the teacher and the student to share a score scale, which is the requirement that Proposition 2 identifies. The price is that OPD inherits the teacher's ordering errors. Consistent with this, the linear RNA teacher is at chance level on held-out patients (mean C-index 0.52, below 0.5 in five of eight cohorts), and OPD with it hurts about as much as with permuted outputs. In the one cohort where the linear teacher is accurate (GBMLGG, 0.756), it improves all four backbones (+0.020 to +0.046, Table 5). We will state this as a limitation and recommend checking a teacher's C-index before distilling from it.

We hope this addresses your concerns, and we are glad to answer any further questions.

Thank you for the additional experiments and clarifications. These strengthen my positive assessment, and I will maintain my original rating.

Reviewer_iqD2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Ordinal Privileged Distillation (OPD) for WSI-only survival prediction. Instead of reconstructing RNA features or matching teacher predictions, OPD transfers pairwise risk ordering from an RNA-only teacher to a WSI student. The paper provides theoretical motivation and evaluates the method on eight cancer datasets and four WSI backbones, showing consistent improvements.

Critical strengths
  1. The motivation is clear and well developed. The paper provides a convincing rationale for transferring task-specific survival information rather than reconstructing molecular features, and for using risk ordering instead of numerical prediction matching.
  2. The proposed method is simple and easy to integrate into existing survival prediction models, without modifying their inference architectures.
  3. The theoretical analyses provide useful support for the motivation. The experiments are also extensive, covering eight datasets and four WSI backbones, with consistently good results and several meaningful ablation studies.
Critical weaknesses
  1. Although OPD achieves good performance, its improvement over Risk MSE, the closest numerical distillation baseline, is relatively modest. More evidence is needed to establish the specific benefit of ordinal supervision.
  2. Some additional experiments and implementation details would help strengthen the conclusions, particularly regarding the quality of the teacher, the reliability of pairwise comparisons, and whether the gains come specifically from privileged information rather than ranking supervision itself.
Questions for authors
  1. The RNA-only teacher plays a central role in OPD, but its own survival prediction performance is not reported. Could the authors provide the C-index of the different teachers, particularly the RNA-only and multimodal teachers, to better explain their differences in distillation effectiveness?
  2. OPD treats all non-tied patient pairs equally, regardless of the teacher's confidence in their relative ordering. How sensitive is the method to uncertain comparisons between patients with nearly identical teacher risk scores? Would filtering low-confidence pairs improve robustness?
  3. Since OPD relies on pairwise comparisons and batch-wise risk standardization, how is the loss implemented when gradient accumulation is used? Does it operate on the actual microbatch or the effective batch of 64 patients? These choices may lead to different training objectives.
  4. Although OPD consistently achieves higher average C-index than Risk MSE, the improvement is relatively small compared with the reported cross-validation variability. Are these gains statistically significant? Paired statistical tests or confidence intervals for the performance differences would help establish the benefit of ordinal supervision over numerical risk matching.
  5. The paper does not compare against a ranking loss constructed directly from observed survival times and censoring information. Does the improvement come specifically from the privileged ordinal information provided by the RNA teacher, or could similar gains be achieved by adding a censoring-aware ranking loss based directly on observed survival outcomes?
4 comments

We sincerely thank you for the constructive feedback and for your positive assessment of our motivation, the simplicity of the method, and the extensive evaluation. W1 is addressed under Q4, and W2 under Q1, Q2, Q3, and Q5.


Q1. The RNA-only teacher plays a central role in OPD, but its own survival prediction performance is not reported. Could the authors provide the C-index of the different teachers, particularly the RNA-only and multimodal teachers, to better explain their differences in distillation effectiveness?

We report the C-index of the different teachers below. The last column is the mean gain of the students distilled from that teacher.

TeacherBRCACRCGBMLGGHNSCKIRCLUADSKCMUCECMeanStudent ΔC×100\Delta C \times 100
RNA-only (used by OPD)0.654±0.0340.617±0.0540.837±0.0520.587±0.0470.671±0.0240.641±0.0220.638±0.0290.600±0.1270.656+3.99
Multimodal MMP0.661±0.0470.677±0.1080.830±0.0470.602±0.0480.684±0.0250.657±0.0250.633±0.0420.592±0.0570.667+2.51
Linear RNA0.445±0.0820.487±0.0960.756±0.0370.471±0.1040.491±0.0480.511±0.0380.440±0.0500.548±0.2320.519−6.38

Based on these results, we make three observations:

  1. A weak teacher is harmful, while an accurate one helps. The linear teacher is below 0.5 in five cohorts, and distilling from it lowers the student there. In the one cohort where it is accurate (GBMLGG, 0.756), it improves all four backbones (Table 5).
  2. Teacher accuracy alone does not determine the gain. The multimodal teacher is more accurate than the RNA-only teacher (0.667 vs. 0.656) but yields a smaller student gain (+2.51 vs. +3.99). The gain grows as the teacher's input overlaps less with the student's (WSI only, then WSI+RNA, then RNA only), in line with our motivation for a teacher that sees only the auxiliary modality.
  3. Students are not bounded by their teacher. The four OPD students reach 0.666–0.699, against 0.656 for the teacher. The students combine the teacher's ordering with WSI evidence and outcome supervision instead of copying the teacher.

Q2. OPD treats all non-tied patient pairs equally, regardless of the teacher's confidence in their relative ordering. How sensitive is the method to uncertain comparisons between patients with nearly identical teacher risk scores? Would filtering low-confidence pairs improve robustness?

Teacher confidence. OPD needs only a scalar risk score from the teacher. Our RNA teacher is a deterministic point predictor, so it gives no calibrated confidence for a pairwise comparison; obtaining one (for example, from an ensemble or a Bayesian teacher) would add requirements on the teacher that OPD is designed to avoid. Near-tied pairs are also a small share of the supervision: at B=64B = 64, about 3% of the pairs are adjacent in teacher rank and about 15% are within five ranks.

To remove pairs with very short follow-up times and early censoring, we trained OPD only on Harrell-comparable pairs. This drops every pair whose earlier observation is censored, including all censored–censored pairs. Compared with OPD on all pairs, training only on Harrell-comparable pairs improves neither accuracy nor stability:

Model (pairs used)BRCACRCGBMLGGHNSCKIRCLUADSKCMUCECMeanMean std
ABMIL (Harrell pairs)0.665±0.0560.646±0.0820.776±0.0420.644±0.0740.675±0.0780.591±0.0310.621±0.0450.672±0.1920.6610.075
ABMIL (all pairs)0.665±0.0560.640±0.0930.780±0.0340.636±0.0790.676±0.0750.591±0.0290.623±0.0430.713±0.1760.6660.073
TransMIL (Harrell pairs)0.682±0.0510.662±0.0270.775±0.0490.634±0.0460.679±0.0780.606±0.0310.602±0.0350.695±0.1400.6670.057
TransMIL (all pairs)0.683±0.0480.664±0.0240.777±0.0420.622±0.0410.682±0.0760.607±0.0390.607±0.0430.717±0.1490.6700.058
MMP-Histo (Harrell pairs)0.682±0.0710.685±0.0960.783±0.0600.639±0.0740.700±0.0730.633±0.0500.647±0.0780.706±0.1620.6840.083
MMP-Histo (all pairs)0.683±0.0690.699±0.0810.783±0.0330.640±0.0730.713±0.0620.630±0.0570.651±0.0790.796±0.1090.6990.070
PANTHER (Harrell pairs)0.694±0.0940.689±0.0900.781±0.0340.629±0.0450.707±0.0610.617±0.0520.637±0.0290.689±0.2130.6800.077
PANTHER (all pairs)0.688±0.0810.690±0.1040.782±0.0310.631±0.0480.710±0.0540.625±0.0540.648±0.0260.719±0.2130.6870.076

Q3. Since OPD relies on pairwise comparisons and batch-wise risk standardization, how is the loss implemented when gradient accumulation is used? Does it operate on the actual microbatch or the effective batch of 64 patients? These choices may lead to different training objectives.

OPD never uses gradient accumulation. Gradient accumulation was used only for the three baselines (G-HANet, LD-CVAE, DisPro), and the phrase "achieved through gradient accumulation when necessary" in Section 4.1 refers to these baselines. Our four backbones (ABMIL, TransMIL, MMP-Histo, and PANTHER), with and without OPD, are trained with a true batch of 64 patients per optimizer step. Both the batch standardization (Eq. 11) and the pair set PB\mathcal P_B (Eq. 10) are computed over these 64 patients.


Q4. Although OPD consistently achieves higher average C-index than Risk MSE, the improvement is relatively small compared with the reported cross-validation variability. Are these gains statistically significant? Paired statistical tests or confidence intervals for the performance differences would help establish the benefit of ordinal supervision over numerical risk matching.

Why the gain over Risk MSE is small. Our Risk MSE standardizes the student and teacher scores within the batch. For z-scores,

1B∑i(rˉip−rˉig)2=2(1−ρP),\frac{1}{B}\sum_i \left(\bar r_i^p - \bar r_i^g\right)^2 = 2(1-\rho_P),

where ρP\rho_P is the within-batch Pearson correlation between the student and teacher risks. Risk MSE is thus itself a scale- and shift-free relational objective: it asks the student to be linearly correlated with the teacher, which already transfers most of the teacher's ordering. Risk MSE and OPD are therefore two instantiations of the principle in Section 3.3: transfer the teacher's risk relationally instead of matching its values. They sit at two levels of invariance:

  • Standardized Risk MSE (affine invariance): invariant to positive affine transformations arg+ba r^g + b (a>0a > 0) of the teacher's scores, but it still forces the student to match relative score gaps.
  • OPD (monotonic invariance): invariant to any strictly increasing transformation of the teacher's scores. It transfers only the comparison signs qijq_{ij} and is a surrogate for Kendall-τ\tau agreement with the teacher. Relaxing affine agreement to monotonic agreement consistently yields the extra +0.68+0.68 gain (30/2/0 W/T/L).

Statistical significance. We tested whether OPD consistently improves on Risk MSE:

ComparisonΔC×100\Delta C \times 100 (mean ± 95% CI)W/T/LPaired tt-testWilcoxon pp
OPD vs. no distillation+3.99 ± 1.1032/0/0t=7.41t = 7.41, p=2.41×10−8p = 2.41 \times 10^{-8}4.66×10−104.66 \times 10^{-10}
OPD vs. DiSKD+3.17 ± 1.1828/1/3t=5.48t = 5.48, p=5.37×10−6p = 5.37 \times 10^{-6}6.85×10−66.85 \times 10^{-6}
OPD vs. Risk MSE+0.68 ± 0.2230/2/0t=6.30t = 6.30, p=5.25×10−7p = 5.25 \times 10^{-7}1.63×10−61.63 \times 10^{-6}
OPD vs. Risk MSE (ABMIL only)+0.65 ± 0.547/1/0t=2.86t = 2.86, p=0.0245p = 0.02450.0178
OPD vs. Risk MSE (TransMIL only)+0.36 ± 0.287/1/0t=3.04t = 3.04, p=0.0190p = 0.01900.0178
OPD vs. Risk MSE (MMP-Histo only)+1.03 ± 0.608/0/0t=4.02t = 4.02, p=0.0050p = 0.00500.0078
OPD vs. Risk MSE (PANTHER only)+0.67 ± 0.498/0/0t=3.26t = 3.26, p=0.0138p = 0.01380.0078

Across the 32 backbone–cohort settings, OPD outperforms Risk MSE in 30, ties in 2 (at three-decimal precision), and loses in none (p=5.25×10−7p = 5.25 \times 10^{-7}). Even within each individual backbone (n=8n = 8 cohorts), the gain of OPD over Risk MSE is statistically significant (p<0.025p < 0.025 for all four backbones).

Q5. The paper does not compare against a ranking loss constructed directly from observed survival times and censoring information. Does the improvement come specifically from the privileged ordinal information provided by the RNA teacher, or could similar gains be achieved by adding a censoring-aware ranking loss based directly on observed survival outcomes?

We added a teacher-free ranking control that uses the observed survival outcomes:

Loss added to NLLPair labelsABMILTransMILMMP-HistoPANTHERΔC×100\Delta C \times 100W/T/L
None (NLL only)–0.6200.6440.6420.657––
Outcome rankingObserved outcomes0.6380.6450.6690.663+1.3524/0/8
OPDRNA teacher0.6660.6700.6990.687+3.9932/0/0

ΔC\Delta C and wins/ties/losses (W/T/L) are computed against NLL-only training over the 32 backbone–cohort settings, as in Table 2.

The results show that a teacher-free ranking loss recovers about one third of the gain, and the rest requires the RNA teacher's ordering. The improvement is therefore tied to the patient-specific ordering from the auxiliary modality, not to the addition of a metric-aligned ranking term.

Full per-cohort results:

ModelBRCACRCGBMLGGHNSCKIRCLUADSKCMUCECMean
ABMIL (NLL only)0.644±0.0580.557±0.1030.767±0.0440.577±0.0580.654±0.0730.556±0.0400.555±0.0430.647±0.1910.620
ABMIL + outcome ranking0.640±0.0540.607±0.0520.771±0.0430.609±0.0710.679±0.0810.562±0.0400.587±0.0340.653±0.1460.638
ABMIL + OPD0.665±0.0560.640±0.0930.780±0.0340.636±0.0790.676±0.0750.591±0.0290.623±0.0430.713±0.1760.666
TransMIL (NLL only)0.676±0.0600.639±0.0350.764±0.0510.612±0.0390.661±0.0910.579±0.0340.583±0.0520.634±0.1870.644
TransMIL + outcome ranking0.679±0.0550.636±0.0440.768±0.0500.595±0.0360.665±0.0840.593±0.0440.581±0.0480.645±0.1520.645
TransMIL + OPD0.683±0.0480.664±0.0240.777±0.0420.622±0.0410.682±0.0760.607±0.0390.607±0.0430.717±0.1490.670
MMP-Histo (NLL only)0.630±0.0600.674±0.1040.732±0.1030.609±0.0550.626±0.0570.586±0.0550.610±0.0780.666±0.1580.642
MMP-Histo + outcome ranking0.678±0.0860.706±0.0930.768±0.0690.578±0.0600.697±0.0680.642±0.0490.616±0.0770.669±0.1530.669
MMP-Histo + OPD0.683±0.0690.699±0.0810.783±0.0330.640±0.0730.713±0.0620.630±0.0570.651±0.0790.796±0.1090.699
PANTHER (NLL only)0.683±0.1100.680±0.1040.764±0.0310.586±0.0750.683±0.0690.619±0.0480.618±0.0640.623±0.1090.657
PANTHER + outcome ranking0.689±0.1050.662±0.0870.775±0.0400.576±0.0620.715±0.0550.617±0.0590.624±0.0500.644±0.1700.663
PANTHER + OPD0.688±0.0810.690±0.1040.782±0.0310.631±0.0480.710±0.0540.625±0.0540.648±0.0260.719±0.2130.687

We hope this addresses your concerns, and we are glad to answer any further questions.

Thank you for the detailed response and additional experiments. The authors have addressed my concerns well. I'll keep my positive score.

Comments on the paper

2 comments
Area_Chair_zXUKOpenReview ↗

Reminder – Please engage with author rebuttal

Dear Reviewers,

We are now approaching the end of the discussion phase. The authors have posted their rebuttal, and I encourage all reviewers to engage actively in the discussion.

Please take the time to: i) read the authors' responses carefully; ii) check whether your concerns and questions have been adequately addressed; and iii) provide the authors with further feedback where needed.

Your active participation at this stage is important for reaching a well-informed final decision.

Thank you for your time and effort.

Best regards,

Your AC

We thank the AC and all four reviewers for their time. This comment summarizes the main concerns, and the new evidence added during the rebuttal. Full per-cohort tables are in the individual responses.

The method in brief. Ordinal Privileged Distillation (OPD) changes only the training loss. A frozen RNA-only survival teacher supplies pairwise risk-ordering labels, and the WSI student learns them through a pairwise logistic loss added to the usual survival NLL. The student's architecture and its WSI-only inference are unchanged.

Points of agreement. Reviewers iqD2, fTzn, and dmXv describe the method as simple and purely loss-driven, with inference left unchanged; iqD2 and fTzn also note the extensive evaluation and ablations (eight cohorts, four backbones). Reviewer JjzX notes the consistent risk stratification and the robustness to the distillation weight λ\lambda.

Main concerns and new evidence. Gains are C-index × 100; W/T/L counts wins/ties/losses over the 32 backbone–cohort settings.

ConcernReviewersNew evidenceResult
Does the gain come from the RNA teacher, or from adding a ranking loss aligned with the C-index?iqD2 (Q5), fTzn (W1)Teacher-free ranking loss on observed outcomes+1.35 (24/0/8) vs. +3.99 (32/0/0) for OPD. The control recovers about one third of the gain.
Is the benefit of ordinal supervision over Risk MSE reliable?iqD2 (W1, Q4), fTzn (W2)Paired tt-tests and Wilcoxon tests+0.68 ± 0.22 (95% CI), 30/2/0, p=5.25×10−7p = 5.25 \times 10^{-7}; p<0.025p < 0.025 within each backbone.
Are the overall gains statistically significant?JjzX (Q2), fTzn (W2)The same tests against NLL-only training and five existing methods+3.99 ± 1.10, 32/0/0, p=2.41×10−8p = 2.41 \times 10^{-8}. The two best OPD students each win 39 of 40 cohort-level comparisons with existing methods.
How accurate are the teachers?iqD2 (Q1), fTzn (W2)C-index of each teacherRNA-only 0.656, multimodal 0.667, linear 0.519, with student gains of +3.99, +2.51, and −6.38. Teacher accuracy alone does not determine the gain, and a chance-level teacher is harmful.
Are unreliable teacher comparisons harmful?iqD2 (Q2), dmXv (Q2)OPD restricted to Harrell-comparable pairsNo improvement in accuracy or stability over using all pairs.
Sensitivity to batch sizedmXv (W1, Q1), iqD2 (Q3)Retraining at batch sizes 8, 16, and 32The gain stays at 3.99–4.10 points, and OPD results vary by at most 0.5 points.
Generalization beyond TCGAJjzXFour CPTAC cohortsOPD improves all 16 backbone–cohort settings, by 4.6 points on average.

Clarifications.

  • Why the gain over Risk MSE is modest (iqD2, fTzn). Our Risk MSE uses batch-standardized scores, so it equals 2(1−ρP)2(1-\rho_P), where ρP\rho_P is the Pearson correlation between student and teacher risks. It is therefore already a scale- and shift-free relational objective, and it accounts for 3.31 of the 3.99 points. Relaxing affine agreement to ordinal agreement adds a further 0.68 points, consistently across backbones.
  • Scope of the theory (fTzn W3, JjzX Q1). Propositions 1 and 2 motivate the choice of target and objective. We do not claim a guarantee that ranking outperforms score matching, as Appendix A.3 states. OPD inherits the teacher's ordering errors: with a chance-level linear teacher, students lose 6.4 points on average.
  • Implementation (iqD2 Q3, fTzn W1). OPD is trained with a true batch of 64 patients, without gradient accumulation. The teacher's training-set scores are in-sample, and held-out patients never enter teacher or student training.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.