LEE- Mekatronik Mühendisliği-Doktora

Bu koleksiyon için kalıcı URI

http://hdl.handle.net/11527/19387

Gözat

Yazar "Aghabiglou, Amir" ile LEE- Mekatronik Mühendisliği-Doktora'a göz atma
  • Öge
    Image reconstruction with deep learning and applications in MR images
    (Graduate School, 2022-04-22) Aghabiglou, Amir ; Ekşioğlu, Ender Mete ; 518172001 ; Mechatronics Engineering
    In this thesis, in the first step, the novel application of the U-Net structure was considered for the important inverse problem of MRI reconstruction. Deep networks are particularly efficient for the speed-up of the MR image reconstruction process by decreasing the data acquisition time, and they can significantly reduce the aliasing artifacts caused by the undersampling in the k-space. On the first try, it is aimed to develop a novel and efficient unfolding U-Net framework for reconstructing MR images from undersampled k-space data. The new framework should have improved reconstruction performance when compared to competing methodologies. In this step, a novel unfolding framework utilizing the U-Net as a sub-block is being proposed. The introduced U-Net unfolding structure is applied to the magnetic resonance image reconstruction problem. The connection between the unfolding U-Nets is realized in the form of a recently developed projection-based updated data consistency layer. The novel structure is implemented in the PyTorch environment, which is one of the standards for deep learning implementations. The recently created fastMRI dataset which forms an important benchmark for MRI reconstruction is used for training and testing purposes. Despite the many challenges in training rather large networks, novel methodologies have enhanced the capability for having clinical-grade MR image reconstruction in real-time. In recent literature, novel developments have facilitated the utilization of deep networks in various image processing inverse problems. In particular, it has been reported multiple times that the performance of deep networks can be improved by using short connections between layers. In the next step of this thesis, a novel MRI reconstruction method is introduced that utilizes such short connections. The dense connections are used inside densely connected residual blocks. Inside these blocks, the feature maps are concatenated to the subsequent layers. In this way, the extracted information is propagated until the last stage of the block. The efficiency of these densely connected residual blocks was evaluated in MRI reconstruction settings, by augmenting different types of effective deep network models with these blocks in novel structures. The quantitative and qualitative results indicate that this original introduction of the densely connected blocks to the MR image reconstruction problem improves the reconstruction performance significantly. In addition, a novel densely connected residual generative adversarial network (DCR-GAN) is proposed for fast and high-quality reconstruction of MR images. DCR blocks enable the reconstruction network to go deeper by preventing feature loss in the sequential convolutional layers. DCR block concatenates feature maps from multiple steps and gives them as the input to subsequent convolutional layers in a feed-forward manner. In this new model, the DCR block's potential to train relatively deeper structures is utilized to improve quantitative and qualitative reconstruction results in comparison to the other conventional GAN-based models. It can be see from the reconstruction results that the novel DCR-GAN leads to improved reconstruction results without a significant increase in the parameter complexity or run times. The GAN-based structures generally suffer from some limitations. They are slow in convergence and they are unstable during the training step. In this work, these limitations of GAN also was addressed by proposing a new wavelet-based structure. To accomplish this, the wavelet transform packet was incorporated into the GAN structure. The wavelet transform is used in the encoding and decoding steps to create this model. In another word, the downsampling and upsampling layers were replaced with Discrete Wavelet Transform (DWT). DWT is used to replace each pooling process during the contraction phase. As DWT is a reversible package, this downsampling approach guarantees that all information can be retained. DWT can also record both the frequency and position information of feature maps, which will aid in the preservation of fine texture. The inverse wavelet transform is employed in the expansion step to upgrade the size of feature maps. Moreover, recent breakthroughs in this field have inspired us to propose another novel deep unfolding structure for MR image reconstruction. In the last step, the model was trained using not only an iteration of the image itself but also utilizing an updated noise level parameter. The noise level parameter is calculated at each iteration using the error between the network output and the initial zero filling estimate. This new parameter is given as an additional input to the network, and it acts as an evolving regularizer for the image manipulation strength of the network over the unrolling iterations. The introduction of this adaptivity over iterations in the training step also improves the deep models reconstructed image quality in the inference stage. Empirical results indicate that the recommended technique can convergence to better reconstruction results when compared to state-of-the-art unfolding structures devoid of such an adaptive parameter. The introduction of the additional adaptive parameter results in an incremental increase in the parameter complexity, and the required reconstruction times also stand very similar. In this thesis, both quantitative and qualitative results were provided and the proposed model's results were evaluated with cutting-edge techniques in the MR image reconstruction field. Three commonly used evaluation metrics of PSNR, SSIM, and NMSE were used to evaluate simulation results. The statistical differences between developed techniques are investigated using the one-way ANOVA method. Additionally, a t-test is used to specify the major difference between the means of the two proposed structures. Additionally, the robustness of the proposed densely connected residual models was verified by testing them with another dataset type without retraining them. The other dataset differs in size and body tissue type compared to the training dataset. The suggested novel structures in this thesis are improved MR image reconstruction performance compared to state-of-the-art techniques regarding all evaluation metrics. They proved their capacity for reconstructing high-quality images. More importantly, the thesis goal was satisfied regarding the acceleration of MR imaging. The proposed models in this thesis are generally considered to be fast enough to be used even in real-time medical imaging.