Abstract
Deep algorithm unrolling has emerged as a powerful model-based approach to develop deep architectures that combine the interpretability of iterative algorithms with the performance gains of supervised deep learning, especially in cases of sparse optimization. This framework is well-suited to applications in biological imaging, where physics-based models exist to describe the measurement process and the information to be recovered is often highly structured. Here, we review the method of deep unrolling, and show how it improves source localization in several biological imaging settings.
Introduction
Biological imaging is a constant tug-of-war between resolution, speed, and interpretability. To observe structures smaller than the diffraction limit of light—such as microtubules or synapses—researchers rely on sparse recovery. This mathematical process attempts to "undo" the blurring caused by a microscope's optics, known as the Point Spread Function (PSF).
Traditionally, this recovery is performed using iterative algorithms like the Iterative Shrinkage-Thresholding Algorithm (ISTA). While mathematically grounded, these methods are often slow, requiring hundreds of iterations to converge, and require manual tuning of hyperparameters. Conversely, standard "black-box" deep learning is fast but often lacks the physical constraints necessary to ensure biological accuracy. Deep algorithm unrolling bridges this gap.
Methodology: The Unrolling Framework
The core innovation discussed is Deep Algorithm Unrolling, specifically the Learned Iterative Shrinkage-Thresholding Algorithm (LISTA). This approach transforms a classical iterative optimization algorithm into a deep neural network architecture.
In this framework, each iteration of the original algorithm is treated as a single layer in a neural network. Instead of using fixed mathematical constants, the network learns optimal weights and thresholds from training data. Because the architecture is derived from a known physical optimization problem, it requires significantly fewer parameters than a generic Convolutional Neural Network (CNN) and remains physically interpretable. This allows the network to implicitly learn the system's PSF during training, reducing the need for perfect manual calibration.
Applications and Results
The paper reviews the application of unrolling across several modalities, including Single-Molecule Localization Microscopy (SMLM) and Ultrasound Localization Microscopy (ULM).
Key Performance Metrics:
- Speed: In 3D Light Field Microscopy (LFM), the unrolled CISTA-net performed neuron localization 10,000 times faster than the iterative ISTA algorithm.
- Temporal Resolution: In SMLM, the LSPARCOM method achieved high-quality reconstructions using only 25 frames, a fraction of the thousands typically required by standard STORM/PALM methods. This enables the imaging of fast-moving live-cell processes.
- Accuracy: In clinical ultrasound, deep unrolled methods showed significantly higher recall and lower localization error than standard fast iterative solvers (like FISTA), particularly at high microbubble concentrations.
Discussion and Future Directions
By treating physics as a prior, unrolled networks are more robust to changes in signal distribution than generic networks. For instance, an unrolled model can transition from imaging microtubules to mitochondria more effectively because it "understands" it is solving an inverse convolution problem.
However, limitations remain. These networks still depend on high-quality training data and sophisticated simulations to bridge the gap to real-world experimental conditions. Future work aims to extend this framework to combinatorial problems in Imaging Transcriptomics (MERFISH) and synapse detection, potentially revolutionizing how gene expression is mapped within complex tissues.
Summary updated