Search

Like in many other research fields, recent developments in computational imaging have focused on developing machine learning (ML) approaches to tackle its main challenges. To improve the performance of computational imaging algorithms, machine learning methods are used for image processing tasks such as noise reduction. Generally, these ML methods heavily rely on the availability of high-quality data on which they are trained. This work explores the application of ML methods, specifically convolutional neural networks (CNNs), in the context of noise reduction for computed tomography (CT) imaging. We utilize a large 2D computed tomography dataset for machine learning to carry out for the first time a comprehensive study on the differences between the observed performances of algorithms trained on simulated noisy data and on real-world experimental noisy data. The study compares the performance of two common CNN architectures, U-Net and MSD-Net, that are trained and evaluated on both simulated and experimental noisy data. The results show that while sinogram denoising performed better with simulated noisy data if evaluated in the sinogram domain, the performance did not carry over to the reconstruction domain where training on experimental noisy data shows a higher performance in denoising experimental noisy data. Training the algorithms in an end-to-end fashion from sinogram to reconstruction significantly improved model performance, emphasizing the importance of matching raw measurement data to high-quality CT reconstructions. The study furthermore suggests the need for more sophisticated noise simulation approaches to bridge the gap between simulated and real-world data in CT image denoising applications and gives insights into the challenges and opportunities in leveraging simulated data for machine learning in computational imaging.

Image quality assessment (IQA) is not just indispensable in clinical practice to ensure high standards, but also in the development stage of novel algorithms that operate on medical images with reference data. This paper provides a structured and comprehensive collection of examples where the two most common full reference (FR) image quality measures prove to be unsuitable for the assessment of novel algorithms using different kinds of medical images, including real-world MRI, CT, OCT, X-Ray, digital pathology and photoacoustic imaging data. In particular, the FR-IQA measures PSNR and SSIM are known and tested for working successfully in many natural imaging tasks, but discrepancies in medical scenarios have been noted in the literature. Inconsistencies arising in medical images are not surprising, as they have very different properties than natural images which have not been targeted nor tested in the development of the mentioned measures, and therefore might imply wrong judgement of novel methods for medical images. Therefore, improvement is urgently needed in particular in this era of AI to increase explainability, reproducibility and generalizability in machine learning for medical imaging and beyond. On top of the pitfalls we will provide ideas for future research as well as suggesting guidelines for the usage of FR-IQA measures applied to medical images.

Neural ordinary differential equations (Neural ODEs) propose the idea that a sequence of layers in a neural network is just a discretisation of an ODE, and thus can instead be directly modelled by a parameterised ODE. This idea has had resounding success in the deep learning literature, with direct or indirect influence in many state of the art ideas, such as diffusion models or time dependant models. Recently, a continuous version of the U-net architecture has been proposed, showing increased performance over its discrete counterpart in many imaging applications and wrapped with theoretical guarantees around its performance and robustness. In this work, we explore the use of Neural ODEs for learned inverse problems, in particular with the well-known Learned Primal Dual algorithm, and apply it to computed tomography (CT) reconstruction.

Due to the merit of establishing volumetric data, X-ray computed tomography (XCT) is increasingly used as a non-destructive evaluation technique in the quality control of advanced manufactured parts with complex non-line-of-sight features. However, the cost of measurement time and data storage hampers the adoption of the technique in production lines. Commercial fast XCT utilises X-ray detectors with fast detection capability, which can be expensive and results a large amount of data. This paper discussed a different approach, where fast XCT was realised via the acquisition of a small number of projection images instead of full projection images. An established total variation (TV) algorithm was used to handle the reconstruction. The paper investigates the feasibility of using the TV algorithm in handling a significantly reduced number of projection images for reconstruction. This allows a reduction of measurement time from fifty-two minutes to one minute for a typical industrial XCT system. It also enables a reduction of data size proportionally. A test strategy including both quantitative and qualitative test metrics was considered to evaluate the effectiveness of the reconstruction algorithm. The qualitative evaluation includes both the signal to noise ratio and the contrast to noise ratio. The quantitative evaluation was established using reference samples with different internal and external geometries. Simulation data were used in the assessment considering various influence factors, such as X-ray source property and instrument noise. The results demonstrated the possibility of using advanced reconstruction algorithms in handling XCT measurements with a significantly limited number of projection images for dimensional measurements., Comment: NDT & E International (2023)

Krylov subspace methods are a powerful family of iterative solvers for linear systems of equations, which are commonly used for inverse problems due to their intrinsic regularization properties. Moreover, these methods are naturally suited to solve large-scale problems, as they only require matrix-vector products with the system matrix (and its adjoint) to compute approximate solutions, and they display a very fast convergence. Even if this class of methods has been widely researched and studied in the numerical linear algebra community, its use in applied medical physics and applied engineering is still very limited. e.g. in realistic large-scale Computed Tomography (CT) problems, and more specifically in Cone Beam CT (CBCT). This work attempts to breach this gap by providing a general framework for the most relevant Krylov subspace methods applied to 3D CT problems, including the most well-known Krylov solvers for non-square systems (CGLS, LSQR, LSMR), possibly in combination with Tikhonov regularization, and methods that incorporate total variation (TV) regularization. This is provided within an open source framework: the Tomographic Iterative GPU-based Reconstruction (TIGRE) toolbox, with the idea of promoting accessibility and reproducibility of the results for the algorithms presented. Finally, numerical results in synthetic and real-world 3D CT applications (medical CBCT and {\mu}-CT datasets) are provided to showcase and compare the different Krylov subspace methods presented in the paper, as well as their suitability for different kinds of problems., Comment: submitted

In this perspective, we argue that despite the democratization of powerful tools for data science and machine learning over the last decade, developing the code for a trustworthy and effective data science system (DSS) is getting harder. Perverse incentives and a lack of widespread software engineering (SE) skills are among many root causes we identify that naturally give rise to the current systemic crisis in reproducibility of DSSs. We analyze why SE and building large complex systems is, in general, hard. Based on these insights, we identify how SE addresses those difficulties and how we can apply and generalize SE methods to construct DSSs that are fit for purpose. We advocate two key development philosophies, namely that one should incrementally grow -- not biphasically plan and build -- DSSs, and one should always employ two types of feedback loops during development: one which tests the code's correctness and another that evaluates the code's efficacy.

X-ray tomographic reconstruction typically uses voxel basis functions to represent volumetric images. Due to the structure in voxel basis representations, efficient ray-tracing methods exist allowing fast, GPU accelerated implementations. Tetrahedral mesh basis functions are a valuable alternative to voxel based image representations as they provide flexible, inhomogeneous partitionings which can be used to provide reconstructions with reduced numbers of elements or with arbitrarily fine object surface representations. We thus present a robust parallelizable ray-tracing method for volumetric tetrahedral domains developed specifically for Computed Tomography image reconstruction. Tomographic image reconstruction requires algorithms that are robust to numerical errors in floating point arithmetic whilst typical data sizes encountered in tomography require the algorithm to be parallelisable in GPUs which leads to additional constraints on algorithm choices. Based on these considerations, this article presents numerical solutions to the design of efficient ray-tracing algorithms for the projection and backprojection operations. Initial reconstruction results using CAD data to define a triangulation of the domain demonstrate the advantages of our method and contrast tetrahedral mesh based reconstructions to voxel based methods.

Tomographic image sizes keep increasing over time and while the GPUs that compute the tomographic reconstruction are also increasing in memory size, they are not doing so fast enough to reconstruct the largest datasets. This problem is often solved by reconstructing data in large clusters of GPUs with enough devices to fit the measured X-ray projections and reconstructed volume. Often this requires tens of GPUs, which is a very economically expensive solution. Access to single-node machines designed to reconstruct using just one or a few GPUs is more common in the field, but current software does not allow iterative reconstruction of volumes that do not fit in those GPUs. In this work, we propose a strategy to execute efficiently the required operations for iterative reconstruction for arbitrarily large images with any number of GPUs with arbitrarily small memories in a single node. Strategies for both the forward and backprojection operators are presented, along with two regularization approaches that are easily generalized to other projection types or regularizers. The proposed improvement also accelerates reconstruction of smaller images on single or multiple GPUs, providing faster code for time-critical medical applications. The resulting algorithm has been added to the TIGRE toolbox, a repository for iterative reconstruction algorithms for general CT, but this memory-saving and problem-splitting strategy can be easily adapted for any other GPU-based CT code., Comment: 18 pages, 11 figures, journal

Iterative algorithms have many advantages for linear tomographic image reconstruction when compared to back-projection based methods. However, iterative methods tend to have significantly higher computational complexity. To overcome this, parallel processing schemes that can utilise several computing nodes are desirable. Popular methods here are row action methods, which update the entire image simultaneously and column action methods, which require access to all measurements at each node. In large scale tomographic reconstruction with limited storage capacity of each node, data communication overheads between nodes becomes a significant performance limiting factor. To reduce this overhead, we proposed a row action method BSGD. The method is based on the stochastic gradient descent method but it does not update the entire image at each iteration, which reduces between node communication. To further increase convergence speeds, an importance sampling strategy is proposed. We compare BSGD to other existing stochastic methods and show its effectiveness and efficiency. Other properties of BSGD are also explored, including its ability to incorporate total variation (TV) regularization and automatic parameter tuning.