FISTA-NET: Deep Algorithm Unrolling for Cerenkov luminescence tomography.

Area of Science:

• Biomedical Imaging
• Medical Physics
• Radiological Sciences

Background:

• Cerenkov luminescence tomography (CLT) offers high sensitivity for 3D radioactive probe imaging in vivo.
• CLT reconstruction accuracy is hindered by simplified models and ill-posed inverse problems.

Purpose of the Study:

• To enhance the performance of CLT reconstruction using a novel model-based deep learning approach.
• To address limitations in current CLT accuracy for improved preclinical imaging.

Main Methods:

• A model-based deep learning network, FISTA-NET, was developed by expanding the Fast Iterative Shrinkage Thresholding Algorithm (FISTA).
• Each network layer represents an iterative step of FISTA, forming a deep neural network.
• Key FISTA parameters were made learnable from training data, optimizing the reconstruction process.

Main Results:

• Numerical simulations demonstrated FISTA-NET's superior positioning and shape recovery capabilities.
• The deep learning approach significantly improved the quality of CLT reconstructions.
• FISTA-NET showed excellent performance in reconstructing the 3D distribution of radioactive probes.

Conclusions:

• FISTA-NET strategy substantially enhances CLT reconstruction quality.
• Improved CLT imaging facilitates better assessment of disease activity and treatment efficacy.
• This deep learning approach holds promise for advancing preclinical molecular imaging.