Brain-metastasis segmentation adapts to new hospitals, modalities and lesion classes: FGCARL reaches 68.28% mean Dice on BraTS-METS 2023
Synopsis
Clinical images arrive in batches by institution, modality and annotated structure, so a segmentation model must absorb a new imaging distribution and new lesion classes without forgetting what it already learned. FGCARL puts bidirectional feature-statistics alignment and Fisher-weighted parameter constraints into one incremental framework that needs no replay of historical patient data. Moving from UCSF-BMSR to BraTS-METS 2023, where domain and classes shift together, FGCARL with Swin UNETR reaches 68.28% mean Dice across the three tumor regions of the new domain, against 60.94% for the strongest comparator, CLAMTS.
Interpretation
A medical image segmentation model can now keep training as both the imaging distribution and the annotated class set change, without holding on to old data. Earlier incremental methods mostly handled either domain shift or class shift, and the two settings were studied separately. In the UCSF-BMSR to BraTS-METS 2023 transfer, the source domain uses three modalities and labels only enhancing tumor, while the target uses four modalities and adds tumor core and surrounding non-enhancing FLAIR hyperintensity; FGCARL with Swin UNETR reaches 68.28% mean Dice on the target domain, against 60.94% for the strongest comparator, CLAMTS.
The method has two parts: one pulls new-domain statistics back toward the old domain at the feature level, the other locks only the parameters that matter for the old task. Mutual Adaptive Instance Representation (MAIR) first applies instance normalization plus a learnable channel-wise affine transform in the forward pass, then restores features to the old domain's recorded mean and standard deviation in the backward pass. Fisher-Guided Knowledge Retention (FGKR) uses diagonal Fisher information to estimate each parameter's influence on the old-domain loss and penalizes its deviation from the old value in proportion; experiments run on Swin UNETR and nnU-Net backbones across six public datasets.
Removing the backward restoration clearly weakens retention of the old domain, so the gain comes from the bidirectional mechanism rather than from swapping in another normalization. With forward MAIR alone, nnU-Net reaches 50.13% Dice on the source domain after adapting from UCSF-BMSR to BrainMetShare. With forward and backward MAIR together, source-domain Dice in the same setting rises to 72.47%, and the target BrainMetShare reaches 66.15%.
Perspective
Evaluation covers six public 3D segmentation datasets, including brain-metastasis MRI and abdominal CT, with nnU-Net and Swin UNETR backbones and Dice percentage as the metric; source and target input channel counts are matched by constructing a T1 subtraction modality.
Source-domain performance still drops after incremental adaptation, for example FGCARL with Swin UNETR falls from 80.13% to 70.63% on UCSF-BMSR; the regularization weight gives a favorable balance near 1.0, and results are reported as Dice.