[Paper Review] DSFormer: A Dual-domain Self-supervised Transformer for Accelerated Multi-contrast MRI Reconstruction

Bo Zhou, Neel Dey+5 more|arXiv (Cornell University)|Jan 26, 2022

Medical Imaging Techniques and Applications27 citations

TL;DR

DSFormer introduces a dual-domain self-supervised transformer for accelerated multi-contrast MRI reconstruction, using deep conditioning and dual-domain losses to match supervised performance.

ABSTRACT

Multi-contrast MRI (MC-MRI) captures multiple complementary imaging modalities to aid in radiological decision-making. Given the need for lowering the time cost of multiple acquisitions, current deep accelerated MRI reconstruction networks focus on exploiting the redundancy between multiple contrasts. However, existing works are largely supervised with paired data and/or prohibitively expensive fully-sampled MRI sequences. Further, reconstruction networks typically rely on convolutional architectures which are limited in their capacity to model long-range interactions and may lead to suboptimal recovery of fine anatomical detail. To these ends, we present a dual-domain self-supervised transformer (DSFormer) for accelerated MC-MRI reconstruction. DSFormer develops a deep conditional cascade transformer (DCCT) consisting of several cascaded Swin transformer reconstruction networks (SwinRN) trained under two deep conditioning strategies to enable MC-MRI information sharing. We further present a dual-domain (image and k-space) self-supervised learning strategy for DCCT to alleviate the costs of acquiring fully sampled training data. DSFormer generates high-fidelity reconstructions which experimentally outperform current fully-supervised baselines. Moreover, we find that DSFormer achieves nearly the same performance when trained either with full supervision or with our proposed dual-domain self-supervision.

Motivation & Objective

• Motivate accelerated multi-contrast MRI (MC-MRI) reconstruction with reduced reliance on fully sampled paired data.
• Develop a transformer-based reconstruction backbone capable of modeling long-range dependencies in MC-MRI data.
• Introduce dual-domain self-supervision (image and k-space) to enable training without fully sampled targets.
• Propose conditioning mechanisms to efficiently share information across contrasts.
• Demonstrate that self-supervised DSFormer can outperform or match fully supervised MC-MRI baselines.

Proposed method

• Propose a deep conditional cascade transformer (DCCT) that cascades Swin Transformer Reconstruction Networks (SwinRN) with data-consistency layers.
• Introduce deep MC-MRI conditioning: K-space Filling (KF) and Channel-wise (CC) conditioning to inject reference contrast information.
• Use a Swin Transformer Block (SwinTB) with Swin Transformer Layers (SwinTL) for non-local feature extraction and reconstruction.
• Train with dual-domain self-supervision: image-domain Appearance Consistency (AC) loss and k-space Partition Data Consistency (PDC) loss.
• AC loss enforces consistency between reconstructions from partitioned k-space inputs in the image domain.
• PDC loss enforces consistency between corresponding partitions in k-space.
• Overall loss combines AC and PDC with balancing weights to guide learning.

Figure 1: An overview of the D ual-domain S elf-supervised Trans Former ( DSFormer ). The Deep Conditional Cascade Transformer (DCCT) is trained in a self-supervised fashion with randomly partitioned undersampled k-space data sets fed into DCCT in parallel. The partition data consistency loss ( $\ma

Experimental results

Research questions

• RQ1 Can a dual-domain self-supervised framework recover high-fidelity MC-MRI reconstructions without fully sampled target data?
• RQ2 Does incorporating MC-MRI conditioning (KF and CC) improve reconstruction quality over single-contrast baselines?
• RQ3 How does DSFormer perform relative to fully supervised MC-MRI methods under various acceleration factors?
• RQ4 What is the impact of the number of cascaded SwinRN blocks on reconstruction quality?

Key findings

• DSFormer achieves state-of-the-art or superior PSNR/SSIM compared to fully supervised MC-MRI baselines across multiple acceleration factors.
• Under T2 target with PD reference, DSFormer reaches PSNR up to 45.05 dB and SSIM up to 0.993 for x2 acceleration in self-supervised mode.
• For x4 acceleration in the T2 target, DSFormer attains 40.31 dB PSNR and 0.985 SSIM (self-supervised), outperforming MCNet and other baselines.
• For x6 acceleration in the T2 target, DSFormer achieves 37.04 dB PSNR and 0.977 SSIM (self-supervised), still outperforming prior supervised methods.
• Across MC-MRI scenarios, DSFormer trained with self-supervision closely matches the performance of fully supervised training, differing by only about 0.03–0.1 dB in PSNR in key cases.
• Ablations show that using both KF and CC conditioning yields the best results, with notable gains over using either alone or neither.

Figure 2: K-space filling (KF) conditioning for $\times 4$ accelerated T2w reconstruction. Undersampled T2w data is combined with fully sampled PD data via KF to generate an initial DCCT input generating much fewer artifacts as compared to zero-padding.