Related Experiment Video
Updated: Jan 8, 2026
Deep Learning-Based Segmentation of Cryo-Electron Tomograms
Published on: November 11, 2022
MSDRN: Multi-scale deep residual network for fluorescence molecular tomography
Xin Zhao1, Liuyuan Zhang1, Chunyu Qiu2
1The Xi'an Key Laboratory of Radiomics and Intelligent Perception, Xi'an, China; School of Information Sciences and Technology, Northwest University, Xi'an, 710127, China.
Background And Objective:
Fluorescence molecular tomography (FMT) is a promising imaging technique that can quantify the internal distribution of tumor in the early stage. However, due to the ill-posed inverse problem caused by the severe photon scattering effect, the promotion of efficiency and accuracy is still an issue for FMT and the reconstruction of the morphological performance is still difficult to meet the practical requirement.
Methods:
In this paper, FMT was employed and a deep-learning-based Multi-Scale Deep Residual Network (MSDRN) is proposed to enhance the reconstruction resolution. During reconstruction, MSDRN maps the measured boundary data into multi-channel feature representations and leverages cascaded residual blocks to deepen the network architecture, thereby enabling comprehensive feature extraction and high-resolution recovery. Specifically, a dual-branch dilated-convolution structure is adopted to enlarge the receptive field, alleviating resolution degradation in multi-source scenarios. A spatial-attention mechanism is further introduced to emphasize the structural similarity of fluorophore distributions. Moreover, an enhanced residual module is designed to accelerate convergence and suppress gradient vanishing. Consequently, MSDRN achieves accurate and high-resolution fluorescent source reconstruction.
Results:
To evaluate the performance of the proposed MSDRN, comprehensive numerical simulations and in-vivo experiments were conducted. The effectiveness of the proposed method is verified in simulation and in-vivo experiments. The results show that the reconstruction accuracy of the proposed method is significantly improved compared with the existing methods, in which Location Error (LE) is reduced by 0.45mm and Dice Similarity Coefficient (Dice) is increased by 42%. The results demonstrate that MSDRN consistently surpasses state-of-the-art approaches in morphological fidelity, localization accuracy, multi-source resolution, and practical in-vivo applicability.
Conclusion:
The proposed MSDRN exhibits superior capabilities in both localizing and recovering the morphological characteristics of fluorescent sources, thereby holding significant potential for advancing the pre-clinical and clinical translation of FMT in early-stage tumor detection.
