SDLFormer: A Sparse and Dense Locality-enhanced Transformer for Accelerated MR Image Reconstruction

TL;DR

SDLFormer introduces a novel transformer architecture with sparse and dense locality enhancements, combining dilated attention and convolution, to improve accelerated MRI image reconstruction in a self-supervised learning framework.

Contribution

It presents a new window-based transformer with dilated attention and convolution, specifically designed for accelerated MRI reconstruction, trained in a self-supervised manner.

Findings

Achieves approximately 1.40 dB higher PSNR than other architectures.

Attains around 0.028 higher SSIM compared to baseline methods.

Demonstrates superior performance across multiple MRI contrasts and undersampling rates.

Abstract

Transformers have emerged as viable alternatives to convolutional neural networks owing to their ability to learn non-local region relationships in the spatial domain. The self-attention mechanism of the transformer enables transformers to capture long-range dependencies in the images, which might be desirable for accelerated MRI image reconstruction as the effect of undersampling is non-local in the image domain. Despite its computational efficiency, the window-based transformers suffer from restricted receptive fields as the dependencies are limited to within the scope of the image windows. We propose a window-based transformer network that integrates dilated attention mechanism and convolution for accelerated MRI image reconstruction. The proposed network consists of dilated and dense neighborhood attention transformers to enhance the distant neighborhood pixel relationship and introduce depth-wise convolutions within the transformer module to learn low-level translation invariant features for accelerated MRI image reconstruction. The proposed model is trained in a self-supervised manner. We perform extensive experiments for multi-coil MRI acceleration for coronal PD, coronal PDFS and axial T2 contrasts with 4x and 5x under-sampling in self-supervised learning based on k-space splitting. We compare our method against other reconstruction architectures and the parallel domain self-supervised learning baseline. Results show that the proposed model exhibits improvement margins of (i) around 1.40 dB in PSNR and around 0.028 in SSIM on average over other architectures (ii) around 1.44 dB in PSNR and around 0.029 in SSIM over parallel domain self-supervised learning. The code is available at https://github.com/rahul-gs-16/sdlformer.git