ESGDiff-FMT: Explicitly sparse guided diffusion for fluorescence molecular tomography

Fluorescence molecular tomography (FMT) is a sensitive optical imaging technology widely used in biomedical research. However, the ill-posed nature of the inverse problem makes FMT reconstruction highly challenging, limiting the quality of the reconstructed images. Although existing iterative optimization and deep learning-based methods can improve reconstruction quality to some extent, the former relies on global regularization and fails to capture local sparse features, while the latter implicitly learns sparsity through network architectures but lacks explicit constraints, limiting the accurate recovery of fluorescent source regions. To address these limitations, we propose an explicitly sparse guided diffusion reconstruction framework, termed ESGDiff-FMT, to enhance the reconstruction of fluorescent sources. First, we propose an asymmetric conditional embedding strategy implemented via a dual-path cross-gating module, which incorporates multi-projection fluorescence signals as stable conditions into the diffusion process and effectively fuses fluorescence signals with reconstructed image features, thereby enhancing their consistency. Second, we introduce the sparsity-aware smooth weighted loss, which adaptively increases the gradient contribution of fluorescence regions while suppressing background interference, explicitly guiding the model to optimize the features of the fluorescent source regions iteratively. Numerical simulations, physical phantom, and in vivo experiments demonstrate that the proposed ESGDiff-FMT achieves superior performance in morphological reconstruction and background noise suppression, even without relying on large-scale datasets.