BBR-Net uses boundary-aware replay to fight forgetting in cardiac segmentation
New method preserves anatomical structure, reducing catastrophic forgetting by focusing on boundaries.
BBR-Net (Boundary-Balanced Replay Network) tackles a key challenge in medical AI: how to keep a model from forgetting old tasks when learning new ones (catastrophic forgetting). Typical replay methods store past images and retrain on them, but they focus on appearance rather than anatomy. The team behind BBR-Net instead prioritizes storing samples that contain rich anatomical boundaries and maintains class balance, ensuring the model retains structurally informative regions.
Tested on cardiac ultrasound segmentation across CAMUS and CardiacNet, BBR-Net matched offline joint-training performance on the source task while adapting well to a new domain. However, in reverse order (starting with noisy CardiacNet), boundary-aware replay failed — because the initial representations were structurally inconsistent. Controlled perturbation experiments confirmed that forgetting scales with structural degradation, not just memory size. This suggests future continual learning systems must assess the reliability of stored anatomical information.
- BBR-Net selects replay samples using boundary-aware priority and class balance to preserve anatomical structure.
- On forward transfer (CAMUS→CardiacNet), BBR-Net matches offline joint-training on source task while reducing forgetting.
- Controlled perturbation shows forgetting increases as structural reliability decreases, not simply as memory capacity limits.
Why It Matters
This reveals that preserving anatomical boundaries, not just appearance, is critical for reliable continual learning in medical imaging.