人工智能
计算机科学
计算机视觉
图像分割
分割
模式识别(心理学)
图像处理
尺度空间分割
图像(数学)
噪音(视频)
特征(语言学)
作者
Guihua Yu,Ke Tan,Tingting Li,Haoran Chen,Daixian Liu,Mengzhu Wang,Hau-Sing So
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2026-04-21
卷期号:: 6541-6545
标识
DOI:10.1109/icassp55912.2026.11463202
摘要
Domain shift limits the clinical deployment of medical image segmentation. Source-free test-time adaptation (TTA) updates models with unlabeled target data during inference, but existing methods remain sensitive to mini-batch statistics and pseudo-label noise, emphasize gradient direction while leaving the step size underspecified, and offer limited pixel-level semantic guidance. We present Magnitude-Aware Semantics-Regularized (MASR), a unified TTA framework that stabilizes optimization and enhances semantic discrimination. MASR introduces gradient magnitude stability (GMS) with a Gradient-Informed Adaptive Learning Rate (GIALR) that modulates step size from gradient-norm fluctuations to complement directional alignment, and employs a lightweight semantic regularizer via a dynamic prototype memory initialized from source statistics and momentum-updated with high-confidence target pixels—without accessing source data. MASR is architecture-agnostic and computationally lightweight. Across heterogeneous medical segmentation benchmarks, MASR achieves state-of-the-art performance with consistent gains over strong TTA baselines; ablations verify the complementary effects of GMS/GIALR and the prototype regularizer.
科研通智能强力驱动
Strongly Powered by AbleSci AI