Unified CT and MRI Pancreas Segmentation for Label-Efficient Cross-Modality Subregion Transfer
Domain-adversarial nnU-Net trained on 4,604 CT/MRI scans achieves 87.31% Dice pancreas segmentation with label-efficient subregion transfer.
A unified 3D pancreas segmentation framework applies domain-adversarial learning to 4,604 heterogeneous CT and MRI scans, aligning CT-MRI features via a latent domain discriminator on a shared nnU-Net encoder-decoder. Whole-pancreas segmentation reaches 87.31% Dice in-distribution and 84.20%-88.09% across external OOD datasets. The transferred encoder achieves 80.53% Dice on MRI and 83.05% on CT for downstream head-body-tail subregion segmentation using only limited MRI subregion annotations.
- 4,604 heterogeneous CT and MRI scans used with domain-adversarial feature alignment
- Whole-pancreas Dice 87.31% in-distribution, 84.20%-88.09% on external OOD sets
- Subregion transfer reaches 80.53% (MRI) and 83.05% (CT) without CT subregion labels
From the original article
1 min read at arxiv.orgRobust medical image segmentation across imaging modalities is challenging because of large differences in appearance and intensity distributions. Models trained on a single modality often show substantial performance drops when applied to unseen domains.
By Ziliang Hong · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.