Learning Task-Specific Sampling Strategy for Sparse-View CT Reconstruction

TL;DR

A deep learning framework that learns task-specific sampling strategies with a multitask approach to train a unified reconstruction network while tailoring optimal sampling strategies for each individual task is proposed.

Abstract

Sparse-View Computed Tomography (SVCT) offers low-dose and fast imaging but suffers from severe artifacts. Optimizing the sampling strategy is an essential approach to improving the imaging quality of SVCT. However, current methods typically optimize a universal sampling strategy for all types of scans, overlooking the fact that the optimal strategy may vary depending on the specific scanning task, whether it involves particular body scans (e.g., chest CT scans) or downstream clinical applications (e.g., disease diagnosis). The optimal strategy for one scanning task may not perform as well when applied to other tasks. To address this problem, we propose a deep learning framework that learns task-specific sampling strategies with a multi-task approach to train a unified reconstruction network while tailoring optimal sampling strategies for each individual task. Thus, a task-specific sampling strategy can be applied for each type of scans to improve the quality of SVCT imaging and further assist in performance of downstream clinical usage. Extensive experiments across different scanning types provide validation for the effectiveness of task-specific sampling strategies in enhancing imaging quality. Experiments involving downstream tasks verify the clinical value of learned sampling strategies, as evidenced by notable improvements in downstream task performance. Furthermore, the utilization of a multi-task framework with a shared reconstruction network facilitates deployment on current imaging devices with switchable task-specific modules, and allows for easily integrate new tasks without retraining the entire model.

Figures (10)

• Figure 1: Limitation of universal sampling strategy in SVCT reconstruction (a) and our proposed task-specific sampling strategy (b). (a) Optimizing a universal sampling strategy for all kinds of CT scans has limitations in finding the optimal strategy for each task. The learned sampling strategy becomes non-optimal when the scanning type changes, thereby undermining the image reconstruction performance. (b) To address this challenge, our proposed task-specific sampling strategy defines different kinds of CT scans as distinct tasks. This approach optimizes strategies for each task within a multi-task framework, aiming to achieve optimal strategies tailored for all types of CT scans.
• Figure 2: The framework of proposed method. The task-specific sampling learns the optimal sampling strategy for each scanning task, while the reconstruction network reconstructs high-quality images from the undersampled projections generated by the sampling strategy. The task-specific sampling strategy learning is performed through a multi-task framework. Undersampled projections are completed into full-view using the sinogram complement network for the following reconstruction. If the task branch contains a downstream-task, an downstream-task network would be trained jointly with reconstruction network.
• Figure 3: Visual results for three types of scans: low-dose non-contrast chest scans, contrast-enhanced CT scans of the abdomen, and non-contrast head CT scans (top to bottom). PSNR/SSIM are marked in yellow at the bottom of each image. All images are reconstructed using 60 views, and regions in red boxes are zoomed below each row. As the arrows show, our method has better performance on details recovering for all three types of scan.
• Figure 4: Violin plots of PSNR (top) and SSIM (bottom) result of sole reconstruction from three tasks. We first conduct the Friedman test to verify if the performances of each method are significant different ($p < 0.05$) over all three tasks. In each task, the Wilcoxon test is conducted between the proposed method and other comparison methods, and $*$ indicates $p < 0.05$.
• Figure 5: Illustration of slice selection for downstream task. We first split the whole volume of one case into $c$ group. In each group, we select one slice randomly during training as the data augmentation and use the middle slice while testing.