MSDRN: Multi-scale deep residual network for fluorescence molecular tomography.

Area of Science:

• Biomedical imaging
• Medical physics
• Computational biology

Background:

• Fluorescence molecular tomography (FMT) is crucial for early-stage tumor detection.
• Severe photon scattering in FMT poses challenges for accuracy and resolution.
• Reconstructing accurate tumor morphology remains difficult for practical FMT applications.

Purpose of the Study:

• To enhance the resolution and accuracy of FMT reconstruction.
• To address the ill-posed inverse problem in FMT caused by photon scattering.
• To improve the morphological performance of FMT for practical requirements.

Main Methods:

• A deep-learning-based Multi-Scale Deep Residual Network (MSDRN) was developed for FMT reconstruction.
• MSDRN utilizes multi-channel feature representations and cascaded residual blocks for comprehensive feature extraction.
• Dual-branch dilated convolutions and a spatial-attention mechanism were employed to improve resolution and emphasize structural similarity.

Main Results:

• MSDRN demonstrated significant improvements in reconstruction accuracy and resolution in simulations and in-vivo experiments.
• Location Error (LE) was reduced by 0.45mm and Dice Similarity Coefficient (Dice) increased by 42% compared to existing methods.
• The method surpassed state-of-the-art approaches in morphological fidelity, localization accuracy, and multi-source resolution.

Conclusions:

• The proposed MSDRN effectively localizes and recovers morphological characteristics of fluorescent sources.
• MSDRN shows significant potential for advancing pre-clinical and clinical translation of FMT.
• This technique holds promise for improved early-stage tumor detection using FMT.