An unsupervised conditional approach to the Generative Latent Optimization framework (cGLO) that benefits from the structural bias of a decoder network and a shared objective between multiple slices to reconstruct sparse-view CT images without requiring a large training dataset.
Abstract
The paper presents a novel unsupervised method called cGLO (Conditional Generative Latent Optimization) for solving Imaging Inverse Problems (IIPs), specifically focusing on the task of sparse-view Computed Tomography (CT) image reconstruction.
Key highlights:
cGLO is built upon the Generative Latent Optimization (GLO) framework and the Deep Image Prior (DIP) approach, combining their advantages.
Unlike supervised deep learning methods, cGLO does not require a fixed experimental setup or a large training dataset. It can be used with or without prior unsupervised training.
cGLO exploits the structural bias of a convolutional decoder network and a shared objective function across multiple slices to reconstruct sparse-view CT images.
Experiments show that cGLO outperforms the current state-of-the-art unsupervised methods, MCG and DIP, in terms of both pixel-wise (PSNR) and structural (SSIM) metrics, especially in scenarios with limited training data and sparse viewing angles.
cGLO is more robust to decreasing training dataset sizes and viewing angles compared to MCG and cSGM (conditional Score-based Generative Models).
The framework developed in this work can be readily applied to other ill-posed IIPs, such as MRI reconstruction, and extended to solve non-linear IIPs.
Conditioning Generative Latent Optimization for Sparse-View CT Image Reconstruction
Stats
"Computed Tomography (CT) is a prominent example of Imaging Inverse Problem (IIP), highlighting the unrivaled performances of data-driven methods in degraded measurements setups like sparse X-ray projections."
"Recent progress have been made using deep learning approaches for CT reconstruction. However, a large part of these novel data-driven methods employ a supervised training setup."
"Current approaches dealing with ill-posed IIPs in an unsupervised way are mostly based on the use of generative models, such as Generative Adversarial Networks (GANs) and Diffusion Models (DMs)."
"DIP uses a randomly initialized U-Net architecture and a fixed input noise. The network weights can then be optimized to solve any ill-posed IIP with known forward model."
Quotes
"Experiments in [41] showed that given sufficient capacity and time/iterations, via gradient descent, the randomly initialized and over-parameterized U-Net can fit the output signal almost perfectly, including the noise η."
"Experiments in [46] and [47] have shown that the GLO approach does not suffer from mode collapse and significantly outperform GANs when trained on small datasets."
"Experiments conducted in this paper cover a wide range of realistic setups with varying amount of available training data and experimental viewing angles."
How can the cGLO framework be extended to solve non-linear Imaging Inverse Problems beyond CT reconstruction
The cGLO framework can be extended to solve non-linear Imaging Inverse Problems (IIPs) beyond CT reconstruction by adapting the optimization process and the network architecture to suit the specific characteristics of the problem. One approach could involve incorporating non-linear activation functions in the decoder network to capture complex relationships in the data. Additionally, the objective function in cGLO can be modified to handle non-linear transformations and constraints. By incorporating non-linearities and adapting the optimization strategy, cGLO can effectively address non-linear IIPs in various imaging modalities.
What are the potential challenges and limitations of applying cGLO to other medical imaging modalities, such as MRI reconstruction
Applying cGLO to other medical imaging modalities, such as MRI reconstruction, may present challenges and limitations due to the unique characteristics of each modality. MRI reconstruction involves different acquisition parameters, noise characteristics, and image artifacts compared to CT imaging. One challenge is the need to tailor the cGLO framework to the specific requirements of MRI data, such as handling complex k-space data and incorporating MRI-specific reconstruction algorithms. Additionally, the computational complexity of MRI reconstruction may pose challenges for real-time applications using cGLO. Ensuring robustness to motion artifacts and optimizing the network architecture for MRI-specific features are also important considerations.
Could the joint learning of reconstruction and segmentation tasks within the cGLO framework lead to further performance improvements
The joint learning of reconstruction and segmentation tasks within the cGLO framework has the potential to lead to further performance improvements in medical imaging applications. By integrating segmentation information into the reconstruction process, cGLO can leverage the complementary nature of these tasks to enhance the overall image quality and accuracy. For example, incorporating segmentation masks as additional constraints in the reconstruction process can improve the delineation of structures and enhance the fidelity of the reconstructed images. Furthermore, the joint learning approach can enable the model to capture spatial relationships between different anatomical structures, leading to more accurate and clinically relevant reconstructions. This integration of reconstruction and segmentation tasks within the cGLO framework has the potential to enhance the overall efficiency and effectiveness of medical imaging workflows.