Abstract
GlobalBioIm is an open-source MATLAB library for solving inverse problems. The library capitalizes on the strong commonalities between forward models to standardize the resolution of a wide range of imaging inverse problems. Endowed with an operator-algebra mechanism, GlobalBioIm allows one to easily solve inverse problems by combining elementary modules in a lego-like fashion. This user-friendly toolbox gives access to cutting-edge reconstruction algorithms, while its high modularity makes it easily extensible to new modalities and novel reconstruction methods. We expect GlobalBioIm to respond to the needs of imaging scientists looking for reliable and easy-to-use computational tools for solving their inverse problems. In this paper, we present in detail the structure and main features of the library. We also illustrate its flexibility with examples from multichannel deconvolution microscopy.
Addressing the Inverse Problem Silo
In fields ranging from astrophysics to medical diagnosis, researchers rarely observe an object directly. Instead, they record indirect, noisy measurements—such as a blurred microscope image or a sparse CT scan—and must mathematically reconstruct the original object. This process is known as an inverse problem.
Historically, software tools for these problems have been "siloed," built specifically for one modality (e.g., X-ray tomography). This lack of modularity forces researchers to rewrite massive amounts of code when experimenting with new regularization techniques or optimization algorithms. GlobalBioIm addresses this gap by providing a unified, object-oriented framework that bridges the distance between elegant mathematical theory and practical implementation.
The GlobalBioIm Architecture
GlobalBioIm is built on four primary abstract classes that serve as the "lego-like" building blocks for any reconstruction task:
- Map: Represents any mathematical operator, whether linear or non-linear.
- LinOp: A specialized subset for linear operators, such as convolutions or Fourier transforms.
- Cost: Functionals to be minimized, typically representing data fidelity or regularization terms.
- Opti: Solvers that implement iterative optimization schemes like ADMM, FISTA, or Primal-Dual.
By overloading standard MATLAB operators, the library allows users to write code that closely mimics mathematical notation. For instance, defining a complex normal operator can be as simple as writing L = H'*H + I, with the library handling the underlying adjoint and composition logic automatically.
Unifying Imaging Modalities
A core strength of GlobalBioIm is its ability to represent diverse imaging modalities through the composition of elementary operators. Whether dealing with Magnetic Resonance Imaging (MRI) or Structured-Illumination Microscopy (SIM), the forward model $\mathcal{H}$ can be expressed as a chain of maps.
This unification allows researchers to swap regularizers or solvers across different modalities with minimal code changes. Moving from Total Variation (TV) to a more complex Hessian-Schatten (HS) norm often requires changing only two lines of code, facilitating rapid benchmarking.
Intelligent Operator Algebra and Efficiency
The library includes a "secret sauce": an automatic simplification engine. When operators are composed, the library intelligently checks for potential simplifications. For example, it recognizes that the adjoint of a convolution is another convolution and can precompute the kernel to save time.
Furthermore, GlobalBioIm offers granular control over memory and speed. Users can toggle flags like doPrecomputation to store intermediate results. In testing, enabling this feature reduced the time to compute a gradient from 1.87 seconds to 0.94 seconds—a 2x speedup achieved without altering the algorithm's logic.
Experimental Validation: Multichannel Deconvolution
The authors demonstrated the library's flexibility using a simulated multichannel deconvolution microscopy task. Using a neuronal culture image, they simulated a 3-channel (RGB) pipeline convolved with Airy disk point-spread functions and corrupted by Gaussian noise (10dB SNR).
Results showed that advanced regularizers like Hessian-Schatten and Good’s Roughness achieved the highest peak SNR (~15dB), significantly outperforming standard Total Variation by mitigating "staircasing" artifacts. The experiment highlighted how easily different solvers (e.g., switching from ADMM to Primal-Dual) could be deployed within the same script.
Conclusion and Future Directions
While GlobalBioIm is currently tied to the MATLAB environment, it supports GPU acceleration to handle large-scale data. It lowers the barrier to entry for state-of-the-art reconstruction, though it still requires a foundational understanding of optimization theory. As an open-source project, it is designed to evolve with the community, incorporating new modules for emerging imaging modalities and non-convex reconstruction methods.
Summary updated