TY - GEN
T1 - Triple Expert Adaptation Networks with Adaptive Prompt Selection for Multi-modal Medical Image Fusion
AU - Wang, Cong
AU - Wang, Kang
AU - Yang, Yang
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.
PY - 2026
Y1 - 2026
N2 - Multi-modal medical image fusion seeks to integrate complementary information from heterogeneous imaging modalities to enhance clinical diagnostics. However, it faces challenges in reconciling channel-wise feature conflicts, spatial misalignments, and redundant feature propagation. In this paper, we propose a Triple Expert Adaptation Network (TEA-Net) to address these limitations through three domain-specific fusion mechanisms: Channel Expert Adaptation Fusion (CEAF), Low-Rank Expert Adaptation Fusion (LoR-EAF), and Spatial Expert Adaptation Fusion (SEAF). The CEAF employs globally dynamic channel attention to resolve intensity conflicts across different modalities. The LoR-EAF focuses on modulating the most relevant interactions between modalities in low-rank to compress extraneous information. The SEAF integrates large-kernel convolutions with cross-modality interactions to align features across different modalities. To boost fusion, we introduce an adaptive prompt selection module that utilizes learnable prompts to guide fusion among experts. Experiments on publicly available datasets demonstrate the superiority of our TEA-Net over state-of-the-art methods. The source codes are available at https://github.com/supersupercong/TEA-Net.
AB - Multi-modal medical image fusion seeks to integrate complementary information from heterogeneous imaging modalities to enhance clinical diagnostics. However, it faces challenges in reconciling channel-wise feature conflicts, spatial misalignments, and redundant feature propagation. In this paper, we propose a Triple Expert Adaptation Network (TEA-Net) to address these limitations through three domain-specific fusion mechanisms: Channel Expert Adaptation Fusion (CEAF), Low-Rank Expert Adaptation Fusion (LoR-EAF), and Spatial Expert Adaptation Fusion (SEAF). The CEAF employs globally dynamic channel attention to resolve intensity conflicts across different modalities. The LoR-EAF focuses on modulating the most relevant interactions between modalities in low-rank to compress extraneous information. The SEAF integrates large-kernel convolutions with cross-modality interactions to align features across different modalities. To boost fusion, we introduce an adaptive prompt selection module that utilizes learnable prompts to guide fusion among experts. Experiments on publicly available datasets demonstrate the superiority of our TEA-Net over state-of-the-art methods. The source codes are available at https://github.com/supersupercong/TEA-Net.
KW - Adaptive Prompt Selection
KW - Channel Expert Adaptation Fusion
KW - Low-rank Expert Adaptation Fusion
KW - Medical Image Fusion
KW - Spatial Expert Adaptation Fusion
UR - https://www.scopus.com/pages/publications/105028274720
U2 - 10.1007/978-3-032-13961-0_14
DO - 10.1007/978-3-032-13961-0_14
M3 - Conference contribution
AN - SCOPUS:105028274720
SN - 9783032139603
T3 - Lecture Notes in Computer Science
SP - 133
EP - 142
BT - Efficient Medical Artificial Intelligence - 1st International Workshop, EMA4MICCAI 2025, Held in Conjunction with MICCAI 2025, Proceedings
A2 - Chen, Tong
A2 - Kim, Jinman
A2 - Wu, Jinge
A2 - Yuan, Kun
A2 - Navab, Nassir
A2 - Xing, Xiaohan
A2 - Du, Yuning
A2 - Padoy, Nicolas
A2 - Ren, Hongliang
A2 - Bai, Long
PB - Springer Science and Business Media Deutschland GmbH
T2 - 1st International Workshop on Efficient Medical Artificial Intelligence, EMA4MICCAI 2025, held in conjunction with MICCAI 2025
Y2 - 23 September 2025 through 23 September 2025
ER -