Compressed Sensing Magnetic Resonance Imaging (CS-MRI) enables the reduction of data acquisition time. Traditional CS-MRI methods are based on iterative modeling an‎d are flexible; however, their execution is often time consuming. Recently, deep learning methods have gained popularity in CS-MRI due to their high performance. However, a key limitation of deep learning approaches is their lack of flexibility in adapting to new data, as they heavily depend on training images an‎d the specific k-space sampling patterns used during training. In this thesis, an Iterative Deep Prior Convolutional Neural Network (IDPCNN) method for MRI image reconstruction is investigated, which combines the advantages of both traditional an‎d deep learning approaches to achieve fast, flexible, an‎d accurate reconstruction. The IDPCNN method consists of two main steps: denoising an‎d image re construction. The denoising step employs a high-level noise reduction technique to smooth the image, while the reconstruction step incorporates prior information from the frequency domain, iteratively enhancing spatial details. The reconstruction quality of IDPCNN surpasses that of the best existing MRI reconstruction methods across various sampling masks an‎d under sampling rates. Its stability, speed, an‎d high-quality reconstructions suggest that IDPCNN holds significant potential for widespread clinical applications.

رضا مختاري

هادي روحاني , رامين جوادي