人工智能
计算机科学
计算机视觉
分割
一致性(知识库)
图像融合
融合
模式识别(心理学)
图像分割
图像(数学)
传感器融合
特征(语言学)
钥匙(锁)
出处
期刊:
日期:2026-01-23
卷期号:: 271-274
标识
DOI:10.1109/iscait69154.2026.11477309
摘要
Multimodal learning leverages complementary information from different modalities to enhance the performance of medical image segmentation. However, current mainstream multimodal learning methods face challenges in medical image segmentation tasks, such as scarcity of annotated data and anatomical misalignments between modalities. To address these issues, this paper proposes a semi-supervised learning framework based on Bidirectional Gated Attention Fusion (BGAF) and multi-level consistency alignment. In terms of fusion mechanism, a bidirectional gated attention fusion module is designed. It extracts modality-invariant knowledge inherent to each modality through bidirectional attention interaction and adaptive gated weight learning across modalities. The extracted information is then integrated into a unified feature fusion layer for deep feature fusion. For consistency constraints, a multi-level constraint system is constructed, which includes feature-level channel semantic alignment and prediction-level contrastive anatomical similarity alignment. By using a channel semantic alignment loss, the framework ensures the alignment of modality-invariant information at the feature level, thereby improving its robustness to misalignments in multimodal scenarios. Additionally, the anatomical structure is regularized by a contrastive anatomical similarity loss, enabling anatomical prediction alignment on unlabeled data in semi-supervised segmentation tasks. Experiments on multimodal medical image datasets show that the proposed method achieves competitive performance compared to other multimodal approaches. Moreover, the method demonstrates outstanding robustness in scenarios involving scarce annotated data and modality misalignments.
科研通智能强力驱动
Strongly Powered by AbleSci AI