B-MIM: Biased Masked Image Modeling for Generalizable Segmentation of Fine-Grained Anatomical Structures
Authors: Sebastián González, Karen Sanchez, José M. Saavedra, Marcelo Pizarro, Bernard Ghanem
Organizations: Universidad de Chile, Santiago, Chile · King Abdullah University of Science and Technology (KAUST), Saudi Arabia · Universidad de los Andes, Santiago, Chile
Self-supervised pretraining enables transferable representations for medical imaging, yet most CT encoders remain biased toward coarse semantic understanding, limiting their sensitivity to fine-grained anatomical structures such as vessels or small tumors. In this paper, we introduce Biased Masked Image Modeling (B-MIM), a modification of the iBOT objective that stochastically reduces global semantic alignment to prioritize local patch reconstruction. This bias encourages the encoder to capture high-frequency morphological details and structural continuity. We curate a multi-institutional CT abdominal dataset of 9,955 filtered studies from 17 public sources and pretrain a 3D Swin Transformer backbone using B-MIM. Across inter-dataset experiments on liver vessel segmentation, the proposed encoder improves topological fidelity (clDice) and achieves competitive Dice scores in tumor segmentation, compared to fully fine-tuned baselines, despite updating only a fraction of the parameters. Our results suggest that reducing global semantic pressure during pretraining enhances generalization to intricate anatomical structures.
Figures & tables
Figure 1: General scheme of the proposed CT encoder and how it is integrated into a downstream segmentation task.
Model
Liver Vessels
Tumors
CRLM
IRCAD
MSD
CRLM
IRCAD
MSD
Dice
clDice
Dice
clDice
Dice
clDice
Dice
Dice
Dice
Swin-nnUNet
0.746
0.780
0.400
0.491
0.598
0.624
0.646
0.572
0.374
VoCo_SSL
0.616
0.645
0.382
0.437
0.488
0.480
0.650
0.369
0.288
VoComni
0.632
0.662
0.455
0.494
0.572
0.588
0.666
0.500
0.423
B-MIM-A
0.744
0.777
0.413
0.500
0.611
0.631
0.631
0.497
0.357
Table 1: Segmentation performance across datasets for different encoder configurations and adaptation strategies. Dice and clDice are reported for liver vessels, while Dice is reported for tumors.
Figure 2: Qualitative comparison for fine-grained liver vessel and tumor segmentation. The top row shows liver vessel predictions, where B-MIM better preserves the continuity of thin vascular branches and reduces local fragmentation compared with Swin-UNETR and VoComni. The bottom row shows tumor segmentation, where B-MIM provides predictions that more closely match the ground truth in the highlighted regions. Zoomed areas emphasize local differences in fine-grained structure delineation.
Medical image segmentation delineates organs, lesions, and other structures in scans to support measurement, diagnosis, treatment planning, and research.