Cross-Cancer Single Domain Generalization

• Cross-Cancer Single Domain Generalization is a learning setting where models trained on a single cancer domain must generalize to multiple unseen cancer types despite shifts in tissue characteristics, modality, and demographics.
• Methodological strategies such as adversarial invariance, modality rebalancing, and data-centric robustness are employed to mitigate domain-specific biases in oncology applications.
• Empirical studies demonstrate that target-free robustness methods, emphasizing diverse multi-cancer training data and careful modality integration, consistently outperform naive, in-domain optimized models.

Searching arXiv for the cited papers and closely related work on cross-cancer single-domain generalization. to=arxiv_search ปมถวายสัตย์ 北京赛车投注query":"Cross-Cancer Single Domain Generalization prognosis (Jiang et al., 11 Jul 2025) FourierDrug (Song et al., 6 Feb 2025) computational pathology domain generalization (Zamanitajeddin et al., 2024) nnU-Net cross cancer PET-CT segmentation (Ghosh et al., 26 Aug 2025) domain adaptation one cancer to another (Cheung et al., 21 Jan 2026)","max_results":10} I’m querying arXiv for the specific IDs and related cross-cancer DG work. Cross-cancer single domain generalization denotes a class of learning problems in which a model is trained without access to target-domain data and is then evaluated on unseen cancer types. In the most explicit formulation, one source cancer domain is used for training and the objective is robust performance on other cancers; closely related target-free cross-cancer settings also appear in pan-cancer drug response prediction, PET-CT tumor segmentation, and computational pathology benchmarks (Jiang et al., 11 Jul 2025, Song et al., 6 Feb 2025, Ghosh et al., 26 Aug 2025, Zamanitajeddin et al., 2024). Across these settings, the central difficulty is domain shift induced by cancer type, tissue morphology, scanner and stain variation, demographics, and modality composition.

1. Definition and relation to adjacent transfer settings

In multimodal prognosis, Cross-Cancer Single Domain Generalization is defined by cancer-type domains $\mathcal D_1,\dots,\mathcal D_K$, where training uses only one source domain $\mathcal D_{\text{train}}=\mathcal D_1$ and testing is performed on unseen target domains $$\mathcal D_{\text{test}}=\{\mathcal D_2,\dots,\mathcal D_K\}$$; the model maps histology and gene-expression inputs to a risk score and is evaluated by concordance index on the unseen cancers (Jiang et al., 11 Jul 2025). In computational pathology more broadly, single-domain generalization is framed as learning $f_\theta:X\to Y$ from one or a few source domains $S\subset D$ such that the expected target risk remains small on an unseen domain $t\notin S$, despite the absence of target data during training (Zamanitajeddin et al., 2024).

A common source of confusion is the boundary between domain generalization and domain adaptation. Domain adaptation explicitly uses unlabeled target-domain samples during training, typically through adversarial alignment. In histopathology, a domain adversarial neural network converts a ResNet50 into a source-target adaptation system by coupling a label predictor with a domain classifier through a gradient reversal layer, and it is trained with labeled source data plus unlabeled target data (Cheung et al., 21 Jan 2026). Cross-cancer single domain generalization, by contrast, excludes such target-domain access during model fitting.

This distinction matters operationally. A model that succeeds under domain adaptation has demonstrated transfer under partial target visibility; a model that succeeds under single-domain generalization has demonstrated target-free robustness. The available literature suggests that these are related but not interchangeable problem statements.

2. Formal problem statements and task-specific instantiations

A general domain-generalization objective in computational pathology writes the empirical source risk as

$$\hat R^s(f)=\frac{1}{n_s}\sum_{i=1}^{n_s}\ell\bigl(f(x_i^s),y_i^s\bigr),$$

the expected target risk as

$$R^t(f)=\mathbb E_{(X,Y)\sim P^t}\bigl[\ell(f(X),Y)\bigr],$$

and optimizes

$$\min_\theta\;\sum_{s\in S}\hat R^s(f_\theta)+\lambda\,\Omega(\theta;S),$$

where $\Omega$ biases the model toward domain-invariant features (Zamanitajeddin et al., 2024). In the prognosis-specific CCSDG formulation, each sample is a triplet $\mathcal D_{\text{train}}=\mathcal D_1$0 of WSI patch features, gene-expression vector, and survival outcome, and the training objective is

$\mathcal D_{\text{train}}=\mathcal D_1$1

with generalization required on the remaining cancer domains (Jiang et al., 11 Jul 2025).

The same structural idea reappears in other cross-cancer tasks. In pan-cancer drug response prediction, source data are GDSC bulk-RNA-seq from $\mathcal D_{\text{train}}=\mathcal D_1$2 cancer types labeled by IC$\mathcal D_{\text{train}}=\mathcal D_1$3-derived sensitive or resistant outcomes; no single-cell or patient data are seen during training, and at test time new expression profiles are passed through the learned encoder, Fourier/LIP projection, and classifier (Song et al., 6 Feb 2025). In PET-CT segmentation, the evaluation is organized around target-only, public-only, and combined training paradigms across oesophageal cancer, AutoPET, and lung cancer cohorts, with robustness measured by Dice Similarity Coefficient, Precision, Recall, and $\mathcal D_{\text{train}}=\mathcal D_1$4 (Ghosh et al., 26 Aug 2025).

The task family is therefore unified less by a single modality than by a shared protocol: train under restricted domain access and test under cross-cancer shift. This suggests that “cross-cancer single domain generalization” functions both as a precise task definition in prognosis and as a broader organizing concept for target-free robustness studies in oncology.

3. Methodological strategies

One major strategy is adversarial invariance. In the drug-response framework "Fourier Asymmetric Attention on Domain Generalization for Pan-Cancer Drug Response Prediction" (Song et al., 6 Feb 2025), the model comprises an encoder $\mathcal D_{\text{train}}=\mathcal D_1$5, a fixed Fourier/LIP projector $\mathcal D_{\text{train}}=\mathcal D_1$6 implemented as a hard-coded matrix $\mathcal D_{\text{train}}=\mathcal D_1$7, a classifier $\mathcal D_{\text{train}}=\mathcal D_1$8, a domain discriminator $\mathcal D_{\text{train}}=\mathcal D_1$9 connected through a gradient-reversal layer, and an asymmetric attention/clustering head. The data flow is

$$\mathcal D_{\text{test}}=\{\mathcal D_2,\dots,\mathcal D_K\}$$0

with $$\mathcal D_{\text{test}}=\{\mathcal D_2,\dots,\mathcal D_K\}$$1 sent simultaneously to the drug-response predictor, the adversarial domain classifier, and the asymmetric clustering objective. The full saddle-point objective is

$$\mathcal D_{\text{test}}=\{\mathcal D_2,\dots,\mathcal D_K\}$$2

The LIP module is interpreted as a truncated discrete Fourier-basis transform, and the asymmetric loss is designed to cluster drug-sensitive samples into a compact group while dispersing resistant samples in the frequency domain.

A second strategy is modality rebalancing plus latent distribution synthesis. In multimodal prognosis, two challenges are identified: degraded weak-modality features and ineffective multimodal fusion (Jiang et al., 11 Jul 2025). The Sparse Dirac Information Rebalancer applies Bernoulli-based sparsification,

$$\mathcal D_{\text{test}}=\{\mathcal D_2,\dots,\mathcal D_K\}$$3

followed by a Dirac-inspired stabilization

$$\mathcal D_{\text{test}}=\{\mathcal D_2,\dots,\mathcal D_K\}$$4

so that weak features do not collapse. The Cancer-aware Distribution Entanglement module then synthesizes a target-like latent distribution by integrating modality-specific Gaussian statistics with a Beta-kernel guidance term, and regularizes the learned joint representation via

$$\mathcal D_{\text{test}}=\{\mathcal D_2,\dots,\mathcal D_K\}$$5

The total loss combines Cox survival terms with this entanglement penalty.

A third strategy is data-centric robustness. The pathology benchmark evaluates 30 domain-generalization algorithms across domain alignment, data augmentation, meta-learning, tailored model design, pretraining, regularization or risk extrapolation, domain separation, and pathology-specific methods (Zamanitajeddin et al., 2024). Its broadest conclusion is methodological rather than architectural: self-supervised learning and stain augmentation consistently outperform many more elaborate DG procedures. In that benchmark, stain-specific interventions are treated as direct countermeasures against a major source of covariate shift in H&E imagery.

Across these studies, the recurring design pattern is not a single canonical architecture but a family of mechanisms for attenuating domain-specific signal: adversarial removal of cancer-type information, explicit rebalancing of weak modalities, latent-space synthesis of unseen distributions, frequency-space decorrelation, and augmentation or pretraining that encourage morphology-driven rather than domain-driven representations.

4. Empirical evidence across oncology tasks

The available evidence spans prognosis, drug response, segmentation, and computational pathology classification. Although the tasks are heterogeneous, each study evaluates robustness under cross-cancer or cross-domain shift rather than purely in-domain accuracy.

Study	Setting	Representative result
(Jiang et al., 11 Jul 2025)	Multimodal prognosis; train on one cancer, test on the other three	SDIR+CADE: average C-index 0.5625; backbone only: 0.5175
(Song et al., 6 Feb 2025)	Drug response; trained solely on in vitro cell line data without target-domain data	Single-cell AUC 0.75–0.82 vs bulk-only MLP 0.62
(Ghosh et al., 26 Aug 2025)	PET-CT segmentation; oesophageal, lung, and AutoPET cohorts	Combined training: DSC 52.9 lung, 40.7 oesophageal, 60.9 AutoPET
(Zamanitajeddin et al., 2024)	Computational pathology benchmark across three tasks	Top average F1: SSL 87.7, StainAug 86.5, ARM 85.5

In multimodal prognosis, the central empirical claim is that multimodal models often generalize worse than unimodal ones in cross-cancer scenarios, and that explicit correction of weak-modality degradation plus entanglement of latent distributions can reverse that pattern (Jiang et al., 11 Jul 2025). On a four-cancer benchmark using TCGA BRCA, BLCA, STAD, and HNSC, the proposed SDIR+CADE model reaches an average C-index of 0.5625 on unseen targets, compared with 0.5370 for TransMIL, 0.5489 for an omics MLP, 0.5175 for SURVPATH, 0.5300 for MixStyle+multimodal, and 0.5369 for DFQ+multimodal. Ablation further shows backbone only at 0.5175, +SDIR at 0.5479, +CADE at 0.5403, and +SDIR+CADE at 0.5625. Kaplan–Meier curves show significant separation with $$\mathcal D_{\text{test}}=\{\mathcal D_2,\dots,\mathcal D_K\}$$6 between high- and low-risk predicted groups.

In pan-cancer drug response prediction, FourierDrug is evaluated under leave-one-out by cancer type and on single-cell and patient-level tasks (Song et al., 6 Feb 2025). In bulk RNA-seq across ten held-out cancers, the reported overall AUROC range is 0.85–0.92, with mean AUROC $$\mathcal D_{\text{test}}=\{\mathcal D_2,\dots,\mathcal D_K\}$$7 on NB, LAML, and MESO. Removing LIP drops AUC by 5–8 percentage points across cancers. On single-cell drug response, a three-layer MLP trained on bulk only yields average AUC $$\mathcal D_{\text{test}}=\{\mathcal D_2,\dots,\mathcal D_K\}$$8, SCAD/scDEAL/CODE-AE achieve 0.65–0.72, and FourierDrug reaches 0.75–0.82. On TCGA patient-level prediction for four drugs, the model outperforms all baselines on 3/4 drugs by 5–12 percentage points in AUROC and is near-tied on Cisplatin. Paired $$\mathcal D_{\text{test}}=\{\mathcal D_2,\dots,\mathcal D_K\}$$9-tests against the best baseline give $f_\theta:X\to Y$0 on both single-cell and patient tasks.

In PET-CT tumor segmentation, the cross-cancer generalization pattern is especially stark (Ghosh et al., 26 Aug 2025). The oesophageal-only 3D nnU-Net obtains the best in-domain DSC on the Australian oesophageal cohort at 57.8, but it collapses on external data, with DSC 1.25 on the Indian lung cohort and 3.4 on AutoPET. AutoPET-only training generalizes more broadly, with DSC 51.6 on lung and 63.5 on AutoPET, but underperforms on oesophageal cancer at 26.7. Combined training produces the most balanced result: DSC 52.9 on lung, 40.7 on oesophageal, and 60.9 on AutoPET, with improved boundary accuracy such as lung $f_\theta:X\to Y$1 reduced from 105.7 mm to 72.9 mm relative to public-only training.

The computational pathology benchmark provides a wider comparative backdrop (Zamanitajeddin et al., 2024). Across CAMELYON17, MIDOG22, and HISTOPANTUM, the top three methods by average F1 are SSL at 87.7, StainAug at 86.5, and ARM at 85.5; paired $f_\theta:X\to Y$2-tests show SSL outperforming StainAug at $f_\theta:X\to Y$3 on full datasets and at $f_\theta:X\to Y$4 on MIDOG22 alone. These results indicate that cross-cancer robustness can depend as much on feature priors and stain perturbation as on explicit DG objectives.

5. Failure modes, caveats, and recurrent misconceptions

One recurring misconception is that multimodal integration is automatically beneficial for out-of-domain generalization. The prognosis study explicitly reports the opposite: multimodal prognosis models often generalize worse than unimodal ones in cross-cancer scenarios, because stronger modalities dominate weaker ones and direct concatenation fails to cover latent target domains (Jiang et al., 11 Jul 2025). The implication is that multimodal fusion without rebalancing or distributional control can amplify domain sensitivity rather than mitigate it.

A second misconception is that strong in-domain performance is evidence of robustness. The PET-CT study directly contradicts this assumption: the oesophageal-only model achieves mean DSC 57.8 in-domain yet falls to 1.25 on lung and 3.4 on AutoPET, which the authors interpret as severe overfitting (Ghosh et al., 26 Aug 2025). Qualitative analysis of failure cases with DSC $f_\theta:X\to Y$5 further identifies very small lesions, misaligned PET/CT pairings, and spurious segmentation of non-tumor structures such as Barrett’s esophagus.

A third misconception is that simple ensembling resolves cross-cancer shift. In histopathology domain adaptation, single-domain supervised CNNs exceed 98% on their own domain but drop to near chance on others, and leave-one-out ensembles do not improve generalization beyond chance (Cheung et al., 21 Jan 2026). In the same study, the effect of stain normalization is sharply target-dependent: lung accuracy drops from 95.56% to 66.60%, whereas breast rises from 49.22% to 81.29% and colon from 78.48% to 83.36%. Cross-cancer robustness therefore cannot be reduced to a uniform preprocessing recipe.

The literature also contains a methodological tension around adversarial alignment. In domain adaptation, DANN substantially improves performance on unlabeled target domains, reaching 95.56% accuracy for lung when trained on labeled breast and colon data and adapted to unlabeled lung data (Cheung et al., 21 Jan 2026). In the pathology DG benchmark, however, adversarial alignment is described as unstable and slow, with other DG methods introducing loops that slow training by approximately 2×–3× while yielding marginal gains of less than 2% average F1 (Zamanitajeddin et al., 2024). A plausible implication is that adversarial methods are highly protocol-dependent: they can be effective when target samples are available for adaptation, but they are not automatically the first-line solution for target-free generalization.

6. Evaluation practice and future directions

The strongest cross-paper recommendation is data diversity. In PET-CT segmentation, the authors conclude that dataset diversity, particularly multi-demographic, multi-center, and multi-cancer integration, outweighs architectural novelty as the key driver of robust generalization (Ghosh et al., 26 Aug 2025). Their combined training strategy improves robustness across all cohorts, and the study recommends future emphasis on large, diverse, expertly annotated multi-cancer cohorts, human-in-the-loop correction frameworks such as MONAI Label, and continued work on small-lesion sensitivity, misregistration correction, and robust post-processing.

In computational pathology, the corresponding evaluation prescription is leave-one-domain-out cross-validation at the tissue-center or cancer-type level, with both Accuracy and binary F1 reported and statistical significance tested across multiple random seeds (Zamanitajeddin et al., 2024). For scenarios with only one cancer or domain available for training, the benchmark recommends starting with a histology-pretrained encoder, applying aggressive stain augmentation together with generic flips and rotations, then adding light domain alignment such as CORAL or MMD only if domain shift persists. It further recommends reserving regularization-based methods such as VREx or IRM, or meta-DG methods such as MLDG, for cases in which simpler augmentation-based baselines fail.

In histopathology transfer settings, interpretability is also treated as part of robustness assessment. Integrated Gradients applied to DANN-based classifiers highlight clusters of dark, densely stained nuclei associated with malignant adenocarcinomas, and blended attribution overlays are used to verify that the learned features correspond to clinically meaningful morphology rather than nuisance signal (Cheung et al., 21 Jan 2026). This suggests that attribution-based audits can serve as a complementary diagnostic when evaluating whether cross-cancer generalization is mechanistically plausible.

Across prognosis and drug response, future methodological work is likely to continue along two trajectories already visible in the literature: explicit correction of representation imbalance across modalities and explicit shaping of latent spaces to mimic unseen target distributions. The prognosis framework positions SDIR and CADE as plug-and-play modules for practical cross-cancer multimodal prognosis (Jiang et al., 11 Jul 2025), while the drug-response framework uses Fourier-basis projection, adversarial domain removal, and asymmetric clustering to train on cell-line data and test directly on single-cell or patient profiles without target-domain exposure (Song et al., 6 Feb 2025). Taken together, these studies portray cross-cancer single domain generalization not as a single algorithm, but as a research program centered on target-free robustness under oncology-specific domain shift.