Research On Magnetic Resonance Image Reconstruction Algorithms Based On Knowledge Distillation

Magnetic Resonance Imaging(MRI)is one of the most important medical imaging techniques with many advantages such as high contrast and no radiation,and is widely used in clinical practice.However,compared with other medical imaging technologies,the limitations of the equipment hardware make MRI acquisition slower.During a long scanning session,inadvertent movement of the scanner can produce severe motion artifacts,resulting in blurred and distorted images,thus causing loss of important diagnostic information.Deep learning-based MRI reconstruction algorithms usually require a large amount of fully sampled data for training,and the models are too complex with huge number of parameters.Inspired by the application of the Knowledge Distillation(KD)method in the field of natural images,some scholars have applied KD to the field of MRI reconstruction to enhance the performance of the model.Specifically,a large and complex teacher network is instructed to train a small and simple student network so that the student network has comparable performance to the teacher network.In this paper,we propose two models based on knowledge distillation and deep convolutional neural networks to reconstruct images with high quality texture details,reducing the reliance on a large number of fully sampled datasets,for experiments and analysis on brain datasets.1.To address the problems of difficulty in acquiring fully sampled data and imaging quality,this paper proposes a semi-supervised distillation learning model based on Swin Transformer for use in MRI image reconstruction work.Firstly,the use of global and local residuals to extract the overall structure and detailed features of the image respectively allows the model to better capture the global and local information of the image,thus improving the clarity of the reconstructed image.Secondly,during the training process,the present article introduces privilege loss to evaluate the quality of knowledge distillation,in order to improve the learning ability of the student network.Furthermore,this algorithm combines shallow features extracted from convolutional layers and deep features obtained from Swin Transformer modules,achieving excellent performance in MR image reconstruction.Finally,the experimental results show that this model produces reconstructed images that are closer to those of supervised models,and preserves more detailed information than unsupervised models.Moreover,it achieves accurate reconstruction results even without a large fully sampled image dataset.2.To further improve the quality of reconstructed images,the present article proposes a contrastive self-distillation MR image reconstruction method based on wavelet transform.First,a reconstructed network of channel splits needs to be constructed from the teacher network,which will act as a compact student network.Secondly,this article studies the characteristics of the knowledge distillation network from the frequency perspective and analyzes that the student network lacks the ability to generate high-quality high-frequency information,which can lead to blurred details in the reconstructed image.The distillation method first decomposes the images into different frequency bands using discrete wavelet transforms,and then distills only the high-frequency structure and detailed information.Compared with the direct distillation learning approach for the generated teacher images,the student network under this approach can focus more on learning in the high-frequency information.Furthermore,this article introduces a contrastive loss function to improve the reconstruction performance of the student network by pulling similar samples closer and pushing dissimilar samples apart.Finally,the experimental results show that the method effectively improves the reconstructed image quality and enhances the more accurate reconstruction performance of the student network compared with the existing model.