SS-CTML: Self-Supervised Cross-Task Mutual Learning for CT Image Reconstruction

TL;DR

SS-CTML introduces a self-supervised cross-task mutual learning framework for CT image reconstruction that simultaneously improves three related tasks—FVCT, SVCT, and LVCT—without requiring paired normal-dose data. The architecture combines prior neural modules, compensation mechanisms, and dual-domain networks to enforce consistency across sinogram and image domains, guided by mutual learning losses among priors and final reconstructions. Experiments on simulated Mayo data and real LDCT demonstrate promising improvements in artifact suppression and structural preservation, with ablations validating the contribution of each component. The work suggests a viable path toward clinical deployment of LDCT reconstruction methods and lays groundwork for a broader CT reconstruction foundation model.

Abstract

Supervised deep-learning (SDL) techniques with paired training datasets have been widely studied for X-ray computed tomography (CT) image reconstruction. However, due to the difficulties of obtaining paired training datasets in clinical routine, the SDL methods are still away from common uses in clinical practices. In recent years, self-supervised deep-learning (SSDL) techniques have shown great potential for the studies of CT image reconstruction. In this work, we propose a self-supervised cross-task mutual learning (SS-CTML) framework for CT image reconstruction. Specifically, a sparse-view scanned and a limited-view scanned sinogram data are first extracted from a full-view scanned sinogram data, which results in three individual reconstruction tasks, i.e., the full-view CT (FVCT) reconstruction, the sparse-view CT (SVCT) reconstruction, and limited-view CT (LVCT) reconstruction. Then, three neural networks are constructed for the three reconstruction tasks. Considering that the ultimate goals of the three tasks are all to reconstruct high-quality CT images, we therefore construct a set of cross-task mutual learning objectives for the three tasks, in which way, the three neural networks can be self-supervised optimized by learning from each other. Clinical datasets are adopted to evaluate the effectiveness of the proposed framework. Experimental results demonstrate that the SS-CTML framework can obtain promising CT image reconstruction performance in terms of both quantitative and qualitative measurements.

Figures (10)

• Figure 1: Overview of the proposed SS-CTML framework. The mutual learning of cross-task (i.e., low-mAs FVCT, SVCT, and LVCT reconstruction tasks) is utilized to regularize the training process. The proposed network architecture mainly consists of prior neural modules(PNMs), compensation modules and dual-domain neural modules(DDNMs).
• Figure 2: The comparison of different reconstruction tasks. From (a) to (c): low-mAs FVCT, SVCT and LVCT reconstruction tasks, respectively. The first row and second row are the reconstruction results of FBP and cross-task mutual learning in image domain, respectively. The display window is [-120, 280] HU.
• Figure 3: (a) Total loss curves for training and validation; and (b) PSNR curves of the low-mAs FVCT, SVCT, and LVCT reconstruction tasks for training and validation.
• Figure 4: The simulated data results of low-mAs FVCT reconstruction task for different methods with different dose levels. The three rows show quarter-dose, sixth-dose, and eighth-dose results, respectively. From (a) to (f): Reference, FBP, RED-CNN, iRadonMap, Dn-DP, and SS-CTML. The display window is [-260, 340] HU.
• Figure 5: The simulated data results of SVCT reconstruction task for different methods with different dose levels. The exposure view number of SVCT is set to 144. The three rows show quarter-dose, sixth-dose, and eighth-dose results, respectively. From (a) to (f): Reference, FBP, FBPConvNet, iRadonMap, FreeSeed, and SS-CTML. The display window is [-260, 340] HU.