Abstract

Glioblastoma is an aggressive primary brain tumor characterized by poor prognosis and substantial spatial heterogeneity in its histopathologic and microenvironmental features. Identifying distinct tumor-associated regions, including cellular tumor, necrosis, microvascular proliferation, immune-cell-rich areas, and infiltrative tissue, is important for diagnosis, grading, and characterization of intratumoral heterogeneity. However, conventional machine-learning approaches and frozen pathology foundation models have limited capacity for task-specific adaptation, while reliance on a single pretrained model may introduce representation biases arising from its pretraining data and architecture. To address these limitations, we developed a dual-foundation-model framework that independently adapts UNI2-h and Virchow2 using low-rank adaptation (LoRA), enabling parameter-efficient task specialization while leveraging complementary representations from the two models. We further introduce slide-aware hierarchical sampling to reduce the dominance of slides with large numbers of patches during training, together with targeted data augmentation and regularization. Predictions from the two adapted foundation models are combined through weighted ensemble fusion to further improve classification performance. On the official challenge validation set, the proposed framework achieved a Macro-F1 of 0.706, an MCC of 0.880, and a composite score of 0.793, defined as the mean of Macro-F1 and MCC. These results suggest the potential of combining parameter-efficient adaptation of pathology foundation models, slide-aware sampling, and complementary ensemble fusion for glioblastoma histopathologic classification. Our code is available at: https://github.com/SJX3689/Dual-Lora.git.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_Path_003.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=fdl3oA9zUX

BibTex

@InProceedings{ZhuYim_DualLoRA_MICCAISAT2026,
        author = { Zhu, Yiming AND Sun, Jiaxiang AND Sadek, Yusuf Khaled AND Zhu, Simeng},
        title = { { Dual-LoRA Adaptation of Pathology Foundation Models for Glioblastoma Histologic Pattern Classification } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17252},
        month = {pending},
        page = {pending}
}


back to top