MASC-Net : modality-aware skip connection network for adaptive feature selection in high-fidelity medical image translation

Abstract

Medical image-to-image translation seeks to synthesize missing modalities for complementary diagnostic information, but existing methods often fail to preserve fine anatomical details, leading to structural loss or blur. Conventional skip connections in encoder-decoder architectures, although effective in detail preservation, often introduce semantic conflicts between source-domain encoder features and target-domain decoder features, thereby degrading generation quality. To address this issue, we propose the Modality-Aware Skip Connection Network (MASC-Net), which integrates a modality-aware attention module into the skip connections. This module adaptively suppresses redundant, irrelevant, and even potentially misleading encoder features, dynamically adjusting the information flow according to the contextual requirements of the decoder, thereby alleviating semantic inconsistencies. In addition, we design a Dual-Branch Feature Fusion Module, which employs advanced feature extraction techniques to simultaneously capture fine-grained local details and global semantic representations at multiple resolutions. This design not only ensures structural consistency but also effectively emphasizes critical pathological features. Experimental results on the BraTS2023 and MRXFDG datasets demonstrate that our method excels in generating anatomical details and exhibits strong generalizability across similar tasks.