Authors
Xiaoliu Luo, Minxue Xiao, Ting Xie, Mengzhu Wang, Huiqing Qi, Joey Tianyi Zhou, Taiping Zhang, Xu Wang
Published in
IEEE transactions on image processing : a publication of the IEEE Signal Processing Society. Volume PP. Nov 17, 2026. Epub Nov 17, 2026.
Abstract
Accurate biomedical image classification under limited supervision remains challenging due to scarce annotations, subtle inter-class differences, and complex disease semantics. Pretrained vision-language models offer a promising foundation for alleviating data scarcity through knowledge transfer, yet effectively adapting them to low-resource biomedical tasks remains nontrivial. To this end, we propose Multi-View Synergistic Learning (MVSL), a unified framework that jointly considers three complementary perspectives: adaptation paradigms, representation granularity, and disease semantic relationships. Specifically, MVSL adopts tailored adaptation strategies for visual and textual encoders, accommodating their distinct representational characteristics while preserving pretrained knowledge. To capture both holistic image semantics and localized pathological evidence, we introduce multi-granularity contrastive learning that aligns global image representations and local visual features with disease-specific textual embeddings. Furthermore, we develop a Disease Semantic Graph (DSG) that leverages disease descriptions generated by large language models to model class-level semantic relationships and regularize adapted textual representations, thereby complementing discriminative learning and helping mitigate excessive specialization toward observed categories. Through the synergy of these complementary perspectives, MVSL enhances fine-grained discrimination while supporting semantic consistency and generalization to unseen categories. Extensive experiments on 11 public biomedical datasets spanning 9 imaging modalities and 10 anatomical regions demonstrate the effectiveness of MVSL in few-shot classification and base-to-novel generalization. The implementation and detailed experimental results are publicly available at https://github.com/luoxiaoliu/MVSL.
PMID:
43043745
Bibliographic data and abstract were imported from PubMed on 18 Nov 2026.
Read full publication at:
Please sign in
to see all details.
Advertisement
Stats
- Recommendations n/a n/a positive of 0 vote(s)
- Views 6
- Comments 0