Research on MRI Image Super-Resolution Reconstruction Algorithm Based on Information Distillation

In recent years, deep learning has demonstrated powerful capabilities in computer vision tasks and injected new energy into MRI image super-resolution reconstruction algorithms. However, existing MRI super-resolution methods based on 2D and 3D CNNs require interpolation to upsample low-resolution images, which increases the computational burden. To address this, this paper proposes an MRI image super-resolution algorithm based on information distillation, called DASID-SR. The algorithm uses information distillation to reduce network parameters and computational complexity while extracting deep features, and applies sub-pixel convolution for direct reconstruction and upsampling, thereby reducing the model's complexity. The core innovations include constructing a Mixed Information Distillation Module (M-IMDB), introducing a differentiable architecture search algorithm in the deep feature extraction module, and implementing dense connections between modules to enhance feature utilization efficiency. Experimental results show that on the Kirby21 and ALVIN-adult datasets, with a reconstruction factor of 3, the proposed method achieves peak signal-to-noise ratios (PSNR) of 40.21dB and 37.56dB, respectively, outperforming other networks by 0.02dB to 2.68dB and 0.02dB to 1.41dB. Moreover, the model's parameter size is only 0.45MB, demonstrating that the proposed algorithm can generate high-quality MRI images while maintaining a low parameter count.