What to Preserve, Where to Adapt: A Depth-Wise Analysis of Forgetting in Continual Gynecological Image Segmentation
Not provided in the content
Abstract
This paper investigates the challenges of continual learning in gynecological image segmentation, focusing on how different encoder-decoder regions affect performance and forgetting.
Reality Card
Forgetting in encoder-decoder architectures during continual learning is significantly influenced by the depth at which updates occur, with early encoder and late decoder regions being most critical for performance.
Ablating early encoder and late decoder regions leads to the largest performance degradation, indicating uneven dependency across the network hierarchy.
The paper does not specify authors or provide full code and analysis pipelines until acceptance, which may limit reproducibility.
Paper to code
Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.