Optical molecular imaging in clinical settings often faces a challenge: balancing patient safety with image quality. High frame rates and low excitation doses can lead to noisy images, making it crucial to find effective denoising methods. Most current deep learning techniques fall short because they ignore the physical models and frequency information, leading to unpredictable results.

Enter DEQ-UMamba, a model-driven approach that uses a proximal gradient descent technique and learnt spatial-frequency characteristics. Instead of simply throwing data at a neural network, DEQ-UMamba breaks down complex noise structures into statistical distributions. This innovative method makes noise estimation and suppression in fluorescent images more effective.

One of the key issues with unfolding networks is computational limitation. DEQ-UMamba cleverly sidesteps this by training an implicit mapping, which directly differentiates the equilibrium point of the convergent solution. This approach ensures stability and avoids the pitfalls of non-convergent behavior.

Each module of DEQ-UMamba aligns with a step in the iterative optimization process. This structure not only boosts performance but also provides clear interpretability. In other words, you can understand what the network is doing at each stage.

Experiments on both clinical and in vivo datasets show that DEQ-UMamba outperforms the current best alternatives while using fewer parameters. This advancement could pave the way for more cost-effective and high-quality clinical molecular imaging.

Reducing Noise in Optical Molecular Images: A New Deep Learning Approach

Actions