Two-Stage Self-supervised Cycle-Consistency Network for Reconstruction of Thin-Slice MR Images

Abstract

The thick-slice magnetic resonance (MR) images are often structurally blurred in coronal and sagittal views, which causes harm to diagnosis and image post-processing. Deep learning (DL) has shown great potential to reconstruct the high-resolution (HR) thin-slice MR images from those low-resolution (LR) cases, which we refer to as the slice interpolation task in this work. However, since it is generally difficult to sample abundant paired LR-HR MR images, the classical fully supervised DL-based models cannot be effectively trained to get robust performance. To this end, we propose a novel Two-stage Self-supervised Cycle-consistency Network (TSCNet) for MR slice interpolation, in which a twostage self-supervised learning (SSL) strategy is developed for unsupervised DL network training. The paired LR-HR images are synthesized along the sagittal and coronal directions of input LR images for network pretraining in the firststage SSL, and then a cyclic interpolation procedure based on triplet axial slices is designed in the second-stage SSL for further refinement. More training samples with rich contexts along all directions are exploited as guidance to guarantee the improved interpolation performance. Moreover, a new cycle-consistency constraint is proposed to supervise this cyclic procedure, which encourages the network to reconstruct more realistic HR images. The experimental results on a real MRI dataset indicate that TSCNet achieves superior performance over the conventional and other SSL-based algorithms, and obtains competitive qualitative and quantitative results compared with the fully supervised algorithm.

Self-supervised super-resolution of 2-D pre-clinical MRI acquisitions

“…Performance across different spatial resolutions The volume-wise peak-signal-to-noise-ratio (PSNR) and structural similarity (SSIM) were calculated for comparative purposes. Results from spline interpolation (order 3) [16][17][18] were used as a baseline benchmark. Following SR/interpolation, a rigid registration was executed utilizing ANTS.…”

Semi‐supervised super‐resolution of diffusion‐weighted images based on multiple references

“…A solution is to sample LR images from the corresponding available HR data by estimating the degradation kernel, 13–16 the estimation accuracy of which severely affects the usability and reliability of the reconstruction results. To address this issue, unsupervised‐learning based SR reconstruction methods have been proposed 17–20 . The most representative models were based on the idea of CycleGAN 21 and DualGAN, 22 which used a forward generator to transform data from domain X to domain Y , and a backward generator to transform data from domain Y back to domain X to achieve cyclic consistency.…”

“…To address this issue, unsupervised-learning based SR reconstruction methods have been proposed. [17][18][19][20] The most representative models were based on the idea of CycleGAN 21 and DualGAN, 22 which used a forward generator to transform data from domain X to domain Y, and a backward generator to transform data from domain Y back to domain X to achieve cyclic consistency. For instance, CinCGAN 23 achieved unsupervised SR by using four generators and two discriminators to construct two CycleGANs.…”

RPLHR-CT Dataset and Transformer Baseline for Volumetric Super-Resolution from CT Scans

“…CNN-based algorithms have achieved outstanding performance in SR for natural images [20] and these techniques have been introduced for volumetric SR [1,4,6,12,13,15,17,18,23,25]. Though significant advances have been made, CNN-based algorithms remain limited by the inherent weaknesses of convolution operators.…”

“…Another impediment to the application of volumetric SR methods is data. Most relevant studies use HR volume as ground truth and degrade it to construct paired pseudo-LR volumes with which to train and evaluate methods [4,15,17,18,23,25]. For instance, Peng et al [17] perform sparse sampling on the depth dimension of thin CT to obtain pseudo thick CT. Zhao et al [25] simulate pseudo-LR MRI by applying an ideal low-pass filter to the isotropic T2-weighted MRI followed by an anti-ringing Fermi filter.…”