Differentiable Neural Architecture Search for Lightweight 3D-MRI Image Super-Resolution Based on Information Distillation

In recent years, the performance of CNN-based image super-resolution reconstruction algorithms has significantly improved; however, most proposed model are handcrafted and thus lack flexibility; Besides, it comes with an increase in the number of parameters and computational complexity. This paper proposes a new method to address the excessive resource consumption of existing algorithms in terms of computational load, memory requirements, and network design. By introducing dense connections between information distillation modules and combining convolutional modules with attention mechanisms, a search space suitable for MRI super-resolution reconstruction is constructed. Using the differentiable architecture search method, this paper optimizes the connections between network modules and attention mechanisms, effectively reducing the number of model parameters while enhancing reconstruction quality. Experimental results demonstrate that the L-DNASR algorithm outperforms traditional non-deep learning methods and state-of-the-art deep learning models on mainstream evaluation metrics such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). In the ANVIL-adults dataset, the reconstruction factor of 2 resulted in a PSNR improvement of 0.19 dB and an SSIM increase of 0.2%. For the reconstruction factor of 3, PSNR improved by 0.15 dB, and SSIM increased by 0.2%, compared to the current state-of-the-art model MBNSR, also 6.67% improvement in Gradient-based Structural Similarity (GSSIM). In addition, the model’s parameter size is only 0.43MB, indicating that the algorithm can generate high-quality MRI images while maintaining a low parameter count.