MR Image Reconstruction from Undersampled Data using Multi Dense Residual Block Network with Spatial and Channel Attention Mechanism and Entropy Loss Function

Magnetic Resonance Imaging (MRI) serves as a leading non-invasive method for medical imaging assessments, finding widespread use in early medical diagnoses due to its exceptional resolution and portrayal of connective softer tissue structures within the human body. Nevertheless, the MRI technique is hindered by an inherently sluggish acquisition procedure. MRI's limitation arises from sequentially acquiring data samples in k-space, with traversal speed constrained by physiological and hardware limitations. Expediting the magnetic resonance imaging (MRI) process can be achieved through the application of compressed sensing (CS) which facilitates the reconstruction of high-quality images from k-space data that has been intentionally undersampled. But the conventional CS-MRI approach is associated with prolonged reconstruction times, and the reconstruction process can introduce artifacts, thereby diminishing the overall quality of the reconstruction. To tackle these challenges, in this research, we devised an enhanced model named Multi Dense Residual Block Network with Spatial and Channel Attention Mechanism (MDRBN-SACA), incorporating an entropy function for the proficient reconstruction of MR images. MDRBN optimally utilizes prior image predictions through dense connections among all MDRBs and each MRDB aggregates past predictions and contributes to subsequent MDRBs, ensuring accurate predictions for reconstructing high-quality MR images. Incorporating a spatial and channel attention mechanism into the dense residual network allows for the extraction of details (edges) from the MR image and recreating an image to closely resemble the ground truth image. The assessment of reconstructed image quality among different CNN algorithms utilized different metrics. The outcomes indicate that MDRBNSACA exhibits superior performance across all quantitative metrics compared to other approaches.