Abstract

Joint prediction of lung cancer N stage (lymph node metastasis), M stage (distant metastasis), and radiogenomic attributes from 3D CT is clinically valuable, but partially correlated tasks can conflict within a shared representation and cause negative transfer. We propose a Shared-Private Subspace Decomposition framework that separates task-shared, task-specific, and task-irrelevant orthogonal subspaces via Structured Orthogonal Decoupling (SOD). ISR, an information-inspired subspace regularizer, guides this allocation. Specifically, SOD initializes an orthonormal basis from warmup feature covariance and refines it with Cayley rotations, preserving orthogonality during training. ISR combines private-only supervision, cross-covariance minimization, and variational bottleneck compression to keep task-specific subspaces discriminative while suppressing redundant and label-irrelevant variation. Each task is predicted using its shared-private representation. On the 6,374-patient Huaxi cohort, our method achieves 0.8110 mAUC, 0.7998 mAP, and 0.411 Exact Match Ratio (EMR); evaluation on the 1,146-patient Shengjing cohort offers a preliminary assessment of cross-center generalization.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CMMCA2026_004.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to Open Review

Open Review Page: https://openreview.net/forum?id=l2NekMnwxv

BibTex

@InProceedings{CheWei_SharedPrivate_MICCAISAT2026,
        author = { Chen, Weiyi AND Liu, Chengcai AND Sang, Haolin AND Huang, Xingyu AND Wang, Chenghao AND Wu, Yi AND Tian, Jie AND Wang, Shuo},
        title = { { Shared-Private Subspace Decomposition for Multi-Task Lung Cancer Classification in CT } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17278},
        month = {pending},
        page = {pending}
}


back to top