What to Preserve, Where to Adapt: Depth-Wise Structural Analysis for Continual Gynecological Image Segmentation
Saqib, Amal
Saqib, Amal
Citations
Altmetric:
Author
Supervisor
Department
Computer Vision
Embargo End Date
Type
Thesis
Date
2026
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Static training assumes that all data is available at the same time, but this does not hold in real-world clinical settings where new scanners, imaging protocols, and patient cohorts introduce continual shifts over time. As a result, data arrives sequentially, requiring models to learn new tasks while preserving performance on previously learned ones. Retraining from scratch is often impractical due to computational costs and privacy and storage constraints, motivating a shift toward continual learning (CL). However, adapting to new data often leads to degradation in performance on previously learned tasks, a problem known as catastrophic forgetting.
Existing CL methods mitigate catastrophic forgetting through replay, regularization, and distillation, but primarily focus on controlling how much model parameters change during adaptation while largely ignoring where these changes occur within the network. In this thesis, we investigate the hypothesis that catastrophic forgetting depends on where adaptation occurs within the network rather than on the number of parameters that change. Using block-level sensitivity analysis on the U-Net encoder–decoder architecture, we examine how different regions of the network contribute to task performance under sequential training and how restricting adaptation to specific depth regions affects forgetting. Our results reveal a clear depth-wise asymmetry: shallow encoder and late decoder layers are highly sensitive and strongly tied to retention, whereas bottleneck-adjacent regions contribute less to both performance and forgetting. Catastrophic forgetting is determined by where adaptation occurs within the network rather than how much the parameters change. Even small updates in retention-critical regions lead to severe forgetting, whereas larger updates in less sensitive regions have a limited impact. These findings provide insight into the stability–plasticity trade-off in continual segmentation and show that structurally guided adaptation can improve retention without additional computational or memory overhead.
Citation
Saqib, Amal, "What to Preserve, Where to Adapt: Depth-Wise Structural Analysis for Continual Gynecological Image Segmentation," M.S. Thesis, Computer Vision, MBZUAI, 2026.
Source
Conference
Keywords
Stability–Plasticity Trade-off, Continual Learning, Catastrophic Forgetting, Structural Analysis of Deep Networks, Depth-wise Adaptation, Medical Image Segmentation