Abstract

Among 455 deduplicated CARE-Liver cases, we observed that liver-spleen morphology had limited discrimination but remained relatively stable across acquisition domains. DINOv2 features were more discriminative within individual domains, whereas source-trained classifiers suffered marked losses after transfer. Label-defined diagnostic refits remained discriminative within each domain, but the fitted directions differed across domains. Thus frozen deep features may retain useful stage-discriminative information, while classifier-target mismatch may partly explain the transfer loss. We ask whether ranks from a source-trained morphology score can guide a deep-feature ranking without target labels. Pairs of target patients with widely separated morphology ranks estimate a lower-to-higher direction in 16-dimensional target PCA space. Its deep rank is averaged with morphology and then refined by an anchored two-group Gaussian model. In label-masked A/B1 episodes, the full method gained 0.036 AUC over morphology for cirrhosis and 0.071 for substantial fibrosis. The same refinement initialized only by morphology gained 0.028 and 0.052; including the patient-pair-derived deep rank therefore added 0.008 (95% CI, -0.009-0.025) and 0.019 (0.001-0.038). For substantial fibrosis, exact-stage-balanced and S1-S2 AUC improved by 0.084 and 0.081. The patient-pair and Gaussian-implied directions were nearly collinear, indicating similar morphology-associated deep variation.

Post-freeze official label-hidden domain-C evaluation yielded AUC 0.808 for cirrhosis and 0.637 for substantial fibrosis after fixed four-class reconciliation. Aggregate-only feedback makes these descriptive external results. The method did not outperform a source-supported comparator and does not learn calibrated four-stage predictions.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CARE_044.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=v43B8bJeNm

BibTex

@InProceedings{LvBoy_Stable_MICCAISAT2026,
        author = { Lv, Boyan AND Jiang, Zhou AND Wang, Lisheng},
        title = { { Stable Morphology, Unstable Deep Classifiers: Morphology-Guided Ranking in Unlabeled Liver MRI Cohorts } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17273},
        month = {pending},
        page = {pending}
}


back to top