Special Session 159: New Developments in Open-Source Software for Inverse Problems

The TIGRE and LION toolboxes for algorithms in applied tomographic reconstruction

Ander Biguri University of Cambridge England Co-Author(s):    Ander Biguri

Abstract: TIGRE (Tomographic Iterative GPU-based Reconstruction Toolbox) and LION (Learning-based Iterative Optimized Networks) are two advanced software toolboxes designed for applied computed tomography (CT) reconstruction, each reflecting a different paradigm in modern imaging. TIGRE Toolbox is an open-source, GPU-accelerated platform developed to support fast and flexible implementation of iterative reconstruction algorithms, particularly for cone-beam CT. It provides a wide range of analytical and iterative methods, enabling researchers to model system geometry accurately and experiment with techniques such as total variation minimization and algebraic reconstruction. Its emphasis on computational efficiency makes it well suited for large-scale 3D problems and real-time applications. In contrast, LION Toolbox focuses on integrating deep learning into the reconstruction pipeline. It enables the design and training of neural networks that mimic or enhance iterative algorithms, often leading to improved image quality and reduced reconstruction time. LION supports hybrid approaches that combine model-based physics with data-driven learning, making it particularly relevant for low-dose and sparse-view CT scenarios. Together, TIGRE and LION illustrate the convergence of classical optimization and modern machine learning, providing complementary tools for advancing research and practical applications in CT reconstruction.

An Advanced Toolbox for Multitask Medical Imaging Consistency (ATOMMIC)

Dimitris Karkalousos Amsterdam University Medical Centers Netherlands Co-Author(s):    Dimitris Karkalousos, Ivana Isgum, Henk A. Marquering, Matthan W.A. Caan

Abstract: The inverse problem of accelerated Magnetic Resonance Imaging (MRI) reconstruction can be solved through Bayesian estimation, where high-quality images are recovered from undersampled measurements in the frequency domain (k-space). Deep learning has driven remarkable progress through iterative optimization schemes, effectively solving the inverse problem end-to-end from acquisition to analysis. Downstream analysis tasks such as quantitative parameter map estimation and segmentation depend directly on reconstruction quality, coupling inverse and inference problems. Existing frameworks typically address these tasks independently, breaking the acquisition and analysis chain and limiting generalization. We present the Advanced Toolbox for Multitask Medical Imaging Consistency (ATOMMIC), an open-source toolbox that unifies AI-driven MRI reconstruction and analysis by natively supporting multitask learning. ATOMMIC enables end-to-end optimization from k-space to task-specific outputs by implementing over 25 deep learning models with harmonized complex-valued and real-valued data support, standardized workflows, and diverse undersampling schemes. Extensive reproducible benchmarks across eight public datasets demonstrate that physics-based models excel at highly accelerated reconstruction and that multitask learning consistently improves performance on both reconstruction and downstream tasks compared to sequential single-task approaches. ATOMMIC provides a comprehensive framework for integrating and evaluating datasets, models, and tasks, and is open-sourced under Apache 2.0 license at \url{https://github.com/wdika/atommic}.

MRpro-An open PyTorch-based MR reconstruction and processing package

Andreas Kofler Physikalisch-Technische Bundesanstalt (PTB), Braunschweig and Berlin Germany Co-Author(s):    Andreas Kofler

Abstract: We present MRpro, an open-source image reconstruction package for MRI built upon PyTorch and open data formats. MRpro consists of three main areas. First, it provides unified datastructures for the processing and handling of MR datasets and their associated metadata, such as sampling trajectories in the Fourier domain. Second, it provides relevant mathematical concepts and operations required to build abstract reconstruction algorithms. Third, to enable the easy development of state-of-the-art learned reconstruction methods, MRpro includes essential building blocks such as so-called data-consistency layers, differentiable optimization layers and well-established neural network backbones. In this talk, we showcase different components of MRpro by considering several learned and non-learned MR imaging problems using both publicly available open-source data, numerical phantoms as well as raw scanner data, demonstrating MRpro`s versatility.

Benchmarking Large Scale Inverse Problems Resolution Algorithms with Benchopt

Benoit Malezieux CNRS France Co-Author(s):

Abstract: Many scientific domains, including 3D tomography and radio astronomy, rely on solving large-scale imaging inverse problems, where a single data sample can reach sizes of 100 million to 1 billion pixels, making both computation and memory critical challenges. This talk introduces a benchmarking framework, built on Benchopt, designed to evaluate the scalability of modern inverse problem solvers on HPC systems. We focus in particular on recent deep learning-based approaches, such as Plug-and-Play methods. The benchmark enables controlled comparisons of reconstruction quality, runtime, and memory usage as computational resources are scaled across distributed architectures. Beyond performance evaluation, this work aims to establish a common ground between methodological innovation and large-scale scientific applications, enabling the development of algorithms that are not only accurate, but also scalable and practically deployable in real-world HPC settings.

The Core Imaging Library: Modular Optimisation for Imaging Inverse Problems

Evangelos Papoutsellis Finden Ltd, University of Manchester England Co-Author(s):

Abstract: The Core Imaging Library (CIL) is an open-source Python framework for solving imaging inverse problems, with a particular emphasis on tomography modalities such as X-ray CT, MRI, and PET. CIL supports multiple stages of the imaging workflow, including data reading, preprocessing, reconstruction and visualisation, making it a versatile platform for developing and testing advanced imaging methods. A central strength of CIL is its modular optimisation framework, which enables users to combine operators, data fidelity terms, and regularisation functionals to formulate and solve a wide range of smooth and non-smooth optimisation problems using first-order methods. Recent developments further extend these capabilities through a flexible stochastic optimisation framework, allowing researchers to configure and test different algorithmic building blocks, sampling strategies, and step-size rules. In this talk, I will introduce CIL, outline its role across the imaging pipeline, and highlight in particular its optimisation module, together with examples from real-world imaging applications that demonstrate its value for developing, testing, and benchmarking advanced methods for inverse problems.

Inversion of Operator Pipelines with Well-Specified Spaces and Backend Heterogeneity

Justus Sagem\\:uller KTH Royal Institute of Technology Sweden Co-Author(s):    Emilien Valat, Ozan \:Oktem, Joakim And\`en

Abstract: The forward mapping for an inverse problem is often best understood in terms of mathematical transformations between certain function spaces. The inversion meanwhile is typically performed with iterative algorithms, looping over calls through a pipeline of operators. This involves arrays as the interchange data type. While arrays are structurally simple, they lack mathematical context and can be ambiguous. Furthermore, one operator may prefer to store its arrays on a GPU, another may only have a CPU implementation, or there may not be enough memory on a single GPU for the entire data size. Working around such limitations requires conversion and copying, which remains bug-prone and detrimental to performance even with help from utilities like DLPack and the Python Array API. The ODL framework wraps arrays into more informative abstractions. This makes it easier to organize code and get it to work correctly. In recent work, we have made these abstractions also capable of tracking different Python libraries for array storage and processing. This gives better control over if and where transfer between backends happens, as well as allowing multiple different devices to simultaneously store parts of the data.

DeepInverse: A Python package for solving imaging inverse problems with deep learning

Jeremy Scanvic LPENSL, France France Co-Author(s):

Abstract: DeepInverse is an open-source PyTorch-based library for imaging inverse problems. DeepInverse implements all steps for image reconstruction, including efficient forward operators, defining and solving variational problems and designing and training advanced neural networks, for a wide set of domains (medical imaging, astronomical imaging, remote sensing, computational photography, compressed sensing and more).

Evolving the ASTRA Toolbox: Computed Tomography in the Age of Deep Learning

Alexander Skorikov CWI Amsterdam Netherlands Co-Author(s):    Alexander Skorikov, Willem Jan Palenstijn

Abstract: The ASTRA Toolbox provides a set of flexible and efficient GPU primitives for transmission-based tomography. It has been powering multiple use cases and software libraries for almost 20 years already, while being continuously maintained, expanded and adapted to the changing demands in the field. One of the important and recent factors for the development of the ASTRA Toolbox has been the tremendously fast growth of Deep Learning adoption in many scientific and engineering applications. In this talk, I will outline what we have been doing to facilitate using the toolbox in combination with Deep Learning libraries and systems, and will provide our outlook for the future direction of its development.

STIR Library: An Open Source Software for Tomographic Image Reconstruction

Charalampos Tsoumpas University of Groningen Netherlands Co-Author(s):    Zekai Li, Daniel Deidda, Nikolaos Efthymiou, Kris Thielemans

Abstract: The Software for Tomographic Image Reconstruction (STIR) library is a long-standing open-source framework for the reconstruction of emission tomography data, with a primary focus on positron emission tomography (PET) and single-photon emission computed tomography (SPECT). This talk presents recent developments in STIR that expand its methodological scope, computational performance, and usability. We highlight advances in iterative and analytic reconstruction algorithms, improved system modelling, support for modern scanner geometries, and enhanced correction techniques, including attenuation, scatter, and motion handling. The integration of GPU acceleration and improved interoperability with Python-based workflows will also be discussed. Emphasis is placed on its modular design, which facilitates reproducible research and rapid prototyping of new reconstruction methods. Through selected application examples, we demonstrate how STIR continues to serve both as a robust reconstruction engine for applied imaging studies and as a flexible research platform for the development and validation of novel inverse problem methodologies in tomographic imaging.

nDTomo: A Python library for X-ray Chemical Imaging and Computed Tomography

Antony Vamvakeros Imperial College London England Co-Author(s):    Antonis Vamvakeros, Evangelos Papoutsellis, Hongyang Dong, Ronan Docherty, Andrew M. Beale, Samuel J. Cooper, Simon D.M. Jacques

Abstract: X-ray chemical imaging techniques, such as XRD-CT, provide vital spatially-resolved information for materials science, yet data processing remains computationally demanding. We present nDTomo, an open-source Python toolkit designed for the simulation, reconstruction, and analysis of hyperspectral tomographic data. Unlike highly abstracted frameworks, nDTomo emphasises pedagogical clarity through a function-centric design, making it ideal for both research and education. The suite integrates standard reconstruction algorithms with deep learning capabilities, including the self-supervised PeakFitCNN for automated peak analysis and GPU-accelerated direct least squares reconstruction (DLSR). We demonstrate the utility of the toolkit through its application in battery research and catalysis, highlighting its ability to correct experimental artifacts such as motor jitter and beam decay. By combining an intuitive PyQt-based graphical user interface (nDTomoGUI) with a modular backend and interactive Jupyter notebooks, nDTomo provides a transparent and reproducible environment for chemical imaging workflows.

OMEGA - Open-source multi-dimensional tomographic reconstruction software for MATLAB, GNU Octave and Python

Ville-Veikko Wettenhovi University of Eastern Finland Finland Co-Author(s):    Niilo Saarlemo

Abstract: In this talk, we present OMEGA - Open-source multi-dimensional tomographic reconstruction software, available for MATLAB, GNU Octave and Python. OMEGA is designed for easy and efficient GPU-based reconstruction of computed tomography (CT), positron emission tomography (PET) and single photon emission computed tomography (SPECT) data, with support for any type of data that uses similar ray-tracing approaches. OMEGA includes dozens of built in algorithms and regularization methods, as well as the ability to perform forward projection ($Ax$) and backprojection ($A^Ty$) operations on any data, in both MATLAB/Octave and Python environments. Unlike other software, OMEGA supports OpenCL, in addition to CUDA, for GPU computations and thus works with any hardware, including CPUs. OMEGA also includes preliminary support for Apple Metal for Mac support. In Python, the forward and backward projections allow interoperability with CuPy, PyTorch, PyOpenCL and ArrayFire data. OMEGA additionally includes several optional features, such as attenuation correction for PET and SPECT, normalization correction for PET, several different projectors for computing the forward and backward projections, methods to prevent out-of-field of view artifacts in CT, several preconditioners, and more. For more information on the software, see the GitHub page: https://github.com/villekf/OMEGA.