Source-linked AI summary
Beyond the Dominant Label: Recovering Coexisting Lesion Evidence with Multiple Instance Learning
Lubin Gan, Jing Zhang, Xinming Zhang, Chengpeng Xu, Siying Wu, Xiaoyan Sun
TL;DR
Weakly supervised MIL typically receives only the most severe diagnosis, although whole-slide images can contain coexisting lesions and standard multiclass training treats the observed label as a complete description. DomSet-MIL aligns censored bag-level supervision with severity-ordered patch routing, and the authors report improved dominant-label prediction and secondary-lesion recovery across natural and controlled settings.
Problem
Dominant-label supervision reports only the most severe finding, while conventional MIL treats that label as fully describing the slide and does not explicitly model why it was selected or localize other present lesions.
Source: §1 · p. 1Method
DomSet-MIL couples censored bag-level likelihood with severity-ordered residual patch routing and top-q pooling to align dominant-label prediction with localized lesion evidence.
Source: §1 · p. 1, §2.2 · p. 2Results
Across natural and controlled settings, DomSet-MIL improves dominant-label prediction and latent or localized secondary-lesion recovery, with ablations supporting complementary roles for censored supervision, residual routing, and sparse pooling.
Source: §4 · p. 5Takeaways & Limitations
The results support modeling the weak-label generation process as a way to obtain richer pathological representations from slide-level annotations.
Source: §4 · p. 5Training observes only the most severe element of an unobserved, nonempty lesion set, rather than the complete set.
Source: §2.1 · p. 2
Abstract
from arXiv · showhide
Whole-slide images may contain coexisting lesions, yet weak supervision typically provides only the most severe diagnosis. We formulate this setting as dominant-label multiple instance learning, where the observed label is the maximum-severity element of a latent lesion set. Standard multiclass MIL treats it as a complete one-hot description and therefore suppresses lower-severity evidence whenever it co-occurs with a dominant lesion. We propose DomSet-MIL, a label-aware framework that aligns the annotation mechanism across bag and instance levels. Dominant-Label Censored Likelihood models the reported diagnosis without declaring unreported lesions absent. Severity-Ordered Residual Routing assigns patch evidence from high to low severity and preserves residual support for secondary findings, while top-$q$ pooling links local routing scores to bag prediction. Experiments across natural and controlled settings show that DomSet-MIL achieves state-of-the-art performance in dominant-label prediction while more effectively recovering latent coexisting-lesion evidence. These findings demonstrate that modeling the dominant-label mechanism can improve diagnostic prediction without sacrificing clinically meaningful secondary findings.
1. INTRODUCTION
Weakly supervised whole-slide classification can report only the most severe lesion, even when a slide contains coexisting findings. DomSet-MIL models this dominant-label mechanism at both bag and patch levels so secondary-lesion evidence is not treated as absent.
1. INTRODUCTION
- Whole-slide images offer comprehensive tissue morphology, but gigapixel resolution makes dense region-level annotation impractical, motivating weak supervision and multiple instance learning.MIL represents a slide as patch embeddings learned from one slide-level diagnosis.Source: §1 · p. 1
- A dominant diagnosis records only the maximum-severity lesion, so standard softmax treats coexisting lower-severity findings as incorrect labels and structural false negatives.Changing the aggregator cannot remove supervision introduced by this annotation rule.Source: §1 · p. 1
- Prior severity-aware MIL still targets the final diagnosis, while multi-label and single-positive-label approaches leave gaps in modeling the annotation mechanism and linking bag-level presence to localized evidence.PAMS targets the final diagnosis; multi-instance multi-label methods assume multiple diagnoses are observed during training, and generic missing-label objectives do not explain label selection.Source: §1 · p. 1
- DomSet-MIL couples Dominant-Label Censored Likelihood (DCL) with Severity-Ordered Residual Routing (SORR) to align latent evidence at bag and patch levels.DCL models the observed diagnosis without declaring unreported lesions absent; SORR routes evidence from higher to lower severity, and top-q pooling links patch assignments to bag scores.Source: §1 · pp. 1–2
- The paper evaluates dominant-label prediction and latent-lesion recovery on BRACS with test-time ROI annotations and on controlled EBRAINS/TCGA mixed-bag benchmarks.Source: §1 · p. 2
2. METHOD
DomSet-MIL models a slide label as the most severe member of an unobserved lesion set, rather than as a complete description of all lesions. Its bag- and patch-level components preserve this severity ordering while retaining evidence for lower-severity findings.
2.1. Dominant-label formulation
- Training observes only the most severe diagnosis in each slide’s nonempty latent lesion set, so a malignant label leaves atypical presence unknown.The framework encodes the three supervision states instead of treating the label as a complete one-hot description.Source: §2.1 · p. 2
2.2. Severity-ordered residual routing
- SORR applies malignant gating first, routes atypical evidence only through residual mass, and treats benign evidence as the parameter-free remainder.Top-q pooling retains a fixed fraction of the strongest route responses to form bag scores.Source: §2.2 · p. 2
2.3. Dominant-label censored likelihood
- DCL models observed labels with a malignant gate and an atypical gate conditional on malignant evidence not being dominant.For malignant slides, the loss is independent of the atypical score, so atypical evidence is censored rather than penalized as a false positive.Source: §2.3 · p. 3
3. EXPERIMENTS
The experiments evaluate dominant-label classification and latent-lesion recovery on natural BRACS slides and controlled mixed-bag tasks. DomSet-MIL leads the reported latent-evidence metrics, while ablations examine the roles of censored supervision, residual routing, and pooling.
3.1. Datasets and evaluation protocol
- BRACS uses a patient-disjoint 395/67/85 train/validation/test split; controlled tasks use known-by-construction latent mixtures, with synthetic class orders for TCGA subtypes.BRACS ROI labels are reserved for test-time evidence analysis; controlled-task training labels are the maximum-severity class while latent sets remain hidden.Source: §3.1 · p. 3
3.3. Latent-evidence recovery
- DomSet-MIL improves over WAN by 0.035–0.038 Set-mAP and 0.089–0.096 Hidden-R across three controlled tasks, and scores 0.625 ROI-mAP and 0.500 Sec-ROI-R on BRACS.The BRACS comparison is against WAN’s 0.594 ROI-mAP and 0.451 Sec-ROI-R.Source: Table 2 · p. 4, §3.3 · p. 4
3.2. Implementation details and baselines
- The baselines include established attention-based methods, recent state-space and severity-aware MIL methods, and matched controls that isolate label modeling from aggregation.All methods use the same frozen UNI features, data splits, and model-selection budget; the matched controls vary patch routing, bag-level supervision, and pooling.Source: §3.2 · p. 4
- On BRACS, DomSet-MIL reaches 0.916 M-AUC, 0.690 BAcc, and 0.071 UER; it leads controlled tasks by 0.004–0.005 M-AUC and 0.009–0.010 BAcc over the strongest baseline.Results are descriptive point estimates, not statistical-significance claims.Source: §3.2 · p. 4
3.4. Ablation study
- Replacing DCL with cross-entropy lowers M-AUC and ROI-mAP by 0.010 and 0.025, raises UER by 0.020, and reduces Sec-ROI-R by 0.067.With DCL retained, replacing SORR with independent gates reduces Sec-ROI-R by 0.038; mean pooling causes the largest M-AUC and ROI-mAP decline among pooling variants.Source: Table 3 · p. 5, §3.4 · p. 5
4. CONCLUSION
The conclusion frames dominant-label supervision as censoring coexisting lower-severity findings and presents DomSet-MIL as a way to retain their evidence. Across natural and controlled settings, the paper reports improved dominant-label prediction and secondary-lesion recovery.
4. CONCLUSION
- Conventional multiclass MIL treats lower-severity lesions censored by a dominant diagnosis as negatives, while DomSet-MIL couples DCL, SORR, and top-q pooling to align bag supervision with patch evidence.The conclusion reports improved dominant-label prediction and latent or localized secondary-lesion recovery across natural and controlled settings.Source: §4 · p. 5
- Future work will examine partially ordered, data-dependent, and uncertain label mechanisms, and whether biomedical knowledge-guided prompts can support lesion-specific evidence learning.Source: §4 · p. 5
5. COMPLIANCE WITH ETHICAL STANDARDS
The study used de-identified or coded pathology images from released resources and involved no new participants, specimens, or animals.
5. COMPLIANCE WITH ETHICAL STANDARDS
- The study used only de-identified or coded images from released BRACS, EBRAINS, and TCGA resources, with no new participants, specimens, or animals.Data use followed applicable access and data-use terms.Source: §5 · p. 6