DomSet-MIL: censored likelihood and residual routing recover coexisting lesion evidence in digital pathology

Modeling dominant-label pathology annotations via right-censored likelihood and severity-ordered residual routing preserves subtle secondary lesion findings suppressed by standard multiclass MIL.

Primary source: arXiv:2610.22091: https://arxiv.org/abs/2610.22091

What’s new: In computational histopathology, gigapixel whole-slide images (WSIs) often contain multiple coexisting tissue lesions spanning benign abnormalities, atypical hyperplasia, and invasive carcinomas. However, standard pathology reporting typically assigns only the single highest-severity diagnostic category to each slide. When weakly supervised multiple instance learning (MIL) algorithms are trained on these archives, standard multiclass formulations treat the reported dominant label as a mutually exclusive one-hot target, which inadvertently penalizes and suppresses patch-level evidence corresponding to secondary, lower-severity lesions. In Beyond the Dominant Label: Recovering Coexisting Lesion Evidence with Multiple Instance Learning (arXiv:2610.22091, 15 Oct 2026), researchers from the University of Science and Technology of China (USTC) and the Hefei Comprehensive National Science Center introduce DomSet-MIL to resolve this structural misalignment.

DomSet-MIL reframes slide-level weak supervision through two complementary mechanisms that bridge bag-level labels and instance-level morphology. First, the authors introduce a Dominant-Label Censored Likelihood loss: rather than assuming non-dominant classes are absent, the objective models the observed slide label as the maximum element of an unobserved lesion set, treating all lower-severity categories as right-censored rather than false. Second, they implement Severity-Ordered Residual Routing across patch representations. Patches are dynamically routed through an ordered hierarchy from highest to lowest severity: high-confidence malignant regions satisfy the dominant diagnostic criterion, while residual patch features are retained and routed downstream to detect coexisting atypical or benign structures. A top-q pooling module then aggregates localized routing scores into calibrated slide predictions. In rigorous evaluations across synthetic benchmarks and real-world clinical breast biopsy cohorts, DomSet-MIL matches or exceeds top-performing MIL baselines (including CLAM, TransMIL, and SemaMIL) in primary slide-level diagnosis, while substantially improving the sensitivity and localization of secondary coexisting lesions without requiring laborious pixel-level segmentations.

Why it matters: Clinical pathology decisions rarely depend on an isolated binary label; identifying precursor lesions, adjacent benign changes, or borderline atypia alongside an invasive tumor is essential for surgical margins, recurrence risk stratification, and personalized oncology protocols. By removing the mutual-exclusivity constraint in standard MIL architectures, DomSet-MIL enables foundation models and clinical diagnostic pipelines to surface full morphological disease spectra directly from existing hospital archives.

Caveats: The framework assumes an ordinal or severity-ordered relationship among disease classes to structure the residual routing hierarchy; applying DomSet-MIL to non-hierarchical, multi-etiology tissue findings (such as concurrent fungal infections and lymphomas) would require generalizing the routing mechanism to partial orders. In addition, evaluation was primarily focused on breast histopathology cohorts where lesion hierarchies are clearly defined under clinical grading guidelines. Validation across broader organ systems and multi-center staining protocols remains an essential next step.