SEQUAL: Self-refining and effective querying active learning with pseudo label divergence score for carotid intima-media segmentation in ultrasound.
Authors
Affiliations (8)
Affiliations (8)
- The UCL Hawkes Institute and the Department of Medical Physics and Biomedical Engineering, University College London, London, UK; School of Mathematical Sciences, Zhejiang University, Hangzhou, China.
- The UCL Hawkes Institute and the Department of Medical Physics and Biomedical Engineering, University College London, London, UK. Electronic address: [email protected].
- Zhejiang Lab, Hangzhou, China.
- Department of Endocrinology, Children's Hospital, Zhejiang University School of Medicine, National Clinical Research Center for Child Health, Hangzhou, China.
- Department of Ultrasound, Children's Hospital, Zhejiang University School of Medicine, National Clinical Research Center for Child Health, Hangzhou, China.
- School of Mathematical Sciences, Zhejiang University, Hangzhou, China.
- Department of Endocrinology, Children's Hospital, Zhejiang University School of Medicine, National Clinical Research Center for Child Health, Hangzhou, China. Electronic address: [email protected].
- Medical Engineering & Engineering Medicine Innovation Center, Hangzhou International Innovation Institute, Beihang University, Hangzhou, China; Key Laboratory of Biomechanics and Mechanobiology (Beihang University), Ministry of Education, Beijing Advanced Innovation Center for Biomedical Engineering, School of Biological and Medical Engineering, Beihang University, Beijing, China; Zhejiang Lab, Hangzhou, China. Electronic address: [email protected].
Abstract
Deep learning has achieved remarkable performance in carotid intima-media (CIM) segmentation from ultrasound images, but its clinical applicability remains limited due to data scarcity, annotation variability, and low image quality. While active learning (AL) aims to minimize labeling cost while model learning, the conventional AL approaches do not fully support sparse and noisy clinical labels that commonly occur in real-world application of CIM segmentation. In this paper, we propose Self-refining and Effective QUerying Active Learning (SEQUAL), a novel AL framework tailored for CIM segmentation under sparse and noisy supervision. SEQUAL introduces a self-refinement mechanism that leverages high-confidence pseudo-labels generated by the model and fuses them with sparse clinical annotations, enabling progressive enhancement of both label quality and model predictions. For efficient and targeted annotation, SEQUAL also proposes a new query strategy based on the Pseudo Label Divergence (PLD) score, which quantifies the information gain introduced by self-refinement. A dual-network design enables fast PLD computation, selecting the most informative samples for annotation and accelerating model convergence. Extensive experiments on carotid ultrasound datasets show that SEQUAL consistently outperforms conventional AL methods in segmentation accuracy, annotation efficiency, and robustness. Moreover, pilot clinical studies demonstrate that CIM measurements derived from SEQUAL's predictions are consistent with expert assessments, confirming its practical utility. The code will be released upon acceptance at https://github.com/yucheng722/SEQUAL.