Theses supervised by Doç. Dr. Tolga Çukur
13 theses · İhsan Doğramacı Bilkent University
Manyetik rezonans görüntüleme için verimli parametre haritalaması
Balanced steady-state free precession (bSSFP) is a magnetic resonance imaging (MRI) sequence that enables high signal-to-noise ratios in short scan times. However, it has elevated sensitivity to main field inhomogeneity, which leads to banding artifacts near regions of relatively large off-resonance shifts. To suppress these artifacts, multiple bSSFP images of the same anatomy are commonly acquired with a set of different RF phase-cycling increments. Joint processing of phase-cycled acquisitions has long been employed to eliminate banding artifacts due to field inhomogeneity. Multiple bSSFP acquisitions can be further used for parameter mapping by exploiting the signal model of phase-cycled bSSFP. While model based approaches for mapping are effective, they often need a large number of acquisitions, inherently limiting scan efficiency. In this thesis, we propose a new method for parameter mapping with improved efficiency and accuracy in phase-cycled bSSFP MRI. The proposed method is based on the elliptical signal model framework for complex bSSFP signals; and introduces an observation about the signal's geometry to the constrained parameter mapping problem, such that the number of unknowns and thereby the required number of acquisitions can be reduced. It also leverages dictionary-based identification to improve estimation accuracy. Simulated, phantom and in vivo experiments demonstrate that the proposed method enables improved parameter mapping with fewer acquisitions.
İşitsel dikkatin dilin beyin üzerindeki gösterimine etkileri
Humans can effortlessly identify target auditory objects during natural listening and shift their focus between different targets. Unique allocation of brain resources would be inefficient for semantic search task. Here, we hypothesize that auditory attention shifts tuning of cortical voxels toward target category and that attention expands the representation of target words while compressing the representation of behaviorally irrelevant words across cortex. To test, we designed an fMRI experiment with a semantic search task. Subjects listened to natural stories twice while searching for words that are semantically related to either 'humans' or 'places'. Fit voxelwise models for two attention tasks were compared to identify semantic tuning shifts in single voxels. Results indicate that attention shifts semantic tuning of single voxels broadly across cortex and attention warps language representation in favor of target words across cortex. We also introduced a novel feature regularization in voxelwise modeling for a naturalistic movie experiment. Feature regularization simply enforces similar model weights over semantically related stimulus features. We tested the proposed method on an fMRI experiment with naturalistic movies. Results suggest that the proposed method offer improved sensitivity in modeling of single voxels. Moreover, we proposed a novel method to improve the sensitivity of phase-sensitive fat-water separation in balanced steady-state free precession (bSSFP) acquisitions. In bSSFP applications using phased-array coils, reconstructed images suffer a lot from spatial sensitivity variations within individual coils. To improve, we first performed region-growing phase correction in individual coil images, then used a linear combination of phase-corrected images. Tests on SSFP angiograms of the thigh, lower leg, and foot suggest that the proposed method enhances fat–water separation in phased-array acquisitions with improved phase estimates.
Değişken yoğunluklu kartezyen olmayan paralel manyetik rezonans görüntülemede hızlı geriçatım
Due to long acquisition times, the use of magnetic resonance imaging (MRI) remains challenging in some applications. Variable-density acquisitions enable scan acceleration while maintaining a desirable trade-off between signal-to-noise ratio (SNR) and spatial resolution. Several image-domain and k-space algorithms were previously proposed for parallel-imaging reconstructions of variable-density acquisitions. However, these methods involve iterative procedures for non-Cartesian data, resulting in substantial computational burden in particular for three-dimensional (3D) reconstructions. An efficient method based on partially parallel imaging with localized sensitivities (PILS) was recently proposed for fast reconstructions of 2D non-Cartesian data. This thesis introduces a generalized image-domain implementation for 3D non-Cartesian variable-density data, and compares it against conventional gridding, PILS, and ESPIRiT (iterative self-consistent parallel imaging reconstruction using eigenvector maps) reconstructions on brain and knee data accelerated at R=2.5 to 4.2. The results indicate that the proposed 3D variable-FOV method outperforms SOS (sum of squares) and PILS methods, and performs equally or better than ESPIRiT reconstruction at less than half of the processing time required by ESPIRiT. Thus, the proposed method provides fast, high-SNR, artifact-suppressed reconstructions.
MRG'de optimizasyon ve makine öğrenimi: Hızlı mr görüntü rekonstrüksiyonu vebeyindeki temsillerin kodlama modellerine uygulanışı
Magnetic Resonance Imaging (MRI) is a non-invasive medical imaging modality that is widely used by clinicians and researchers to picture body anatomy and neuronal function. However, long scan time remains a major problem. Recently, multiple techniques have emerged that reduce the acquired MRI signal samples, hence dramatically accelerating the acquisition. These techniques involve sophisticated signal reconstruction procedures that in essence require solving regularized optimization problems, and clinical adoption of accelerated MRI critically relies on self-tuning solutions for these problems. Further to this, recent experimental approaches in cognitive neuroscience favor employing naturalistic audio-visual stimuli that closely resemble humans' daily-life experience. Yet, these modern paradigms inevitably lead to huge functional MRI (fMRI) datasets that require advanced statistical and computational techniques to uncover the large amount of embedded information. Here, we propose a novel efficient data-driven self-tuning reconstruction method for accelerated MRI. We demonstrate superior performance of the proposed method across various simulated and in vivo datasets and under various scan configurations. Furthermore, we develop statistical analysis tools to investigate the neural representation of hundreds of action categories in natural movies in the brain via fMRI, and study their attentional modulations. Finally, we develop a model-based framework to estimate temporal extent of semantic information integration in the brain, and investigate its attentional modulations using fMRI data recorded during natural story listening. In short, the methodological and analytical approaches introduced in this thesis greatly benefit clinical utility of accelerated MRI, and enhance our understanding of brain function in daily life.
Deep learning for accelerated MR imaging
Magnetic resonance imaging is a non-invasive imaging modality that enables multi-contrast acquisition of an underlying anatomy, thereby supplementing multitude of information for diagnosis. However, prolonged scan duration may prohibit its practical use. Two mainstream frameworks for accelerating MR image acquisitions are reconstruction and synthesis. In reconstruction, acquisitions are accelerated by undersampling in k-space, followed by reconstruction algorithms. Lately deep neural networks have offered significant improvements over traditional methods in MR image reconstruction. However, deep neural networks rely heavily on availability of large datasets which might not be readily available for some applications. Furthermore, a caveat of the reconstruction framework in general is that the performance naturally starts degrading towards higher acceleration factors where fewer data samples are acquired. In the alternative synthesis framework, acquisitions are accelerated by acquiring a subset of desired contrasts, and recovering the missing ones from the acquired ones. Current synthesis methods are primarily based on deep neural networks, which are trained to minimize mean square or absolute loss functions. This can bring about loss of intermediate-to-high spatial frequency content in the recovered images. Furthermore, the synthesis performance in general relies on similarity in relaxation parameters between source and target contrasts, and large dissimilarities can lead to artifactual synthesis or loss of features. Here, we tackle issues associated with reconstruction and synthesis approaches. In reconstruction, the data scarcity issue is addressed by pre-training a network on large readily available datasets, and fine-tuning on just a few samples from target datasets. In synthesis, the loss of intermediate-to-high spatial frequency is catered for by adding adversarial and high-level perceptual losses on top of traditional mean absolute error. Finally, a joint reconstruction and synthesis approach is proposed to mitigate the issues associated with both reconstruction and synthesis approaches in general. Demonstrations on MRI brain datasets of healthy subjects and patients indicate superior performance of the proposed techniques over the current state-of-the art ones.
Hızlandırılmış 3D MRG için Derin Öğrenme
Magnetic resonance imaging (MRI) offers the flexibility to image a given anatomic volume under a multitude of tissue contrasts. Yet, scan time considerations put stringent limits on the quality and diversity of MRI data. The gold-standard approach to alleviate this limitation is to recover high-quality images from data undersampled across various dimensions, most commonly the Fourier domain or contrast sets. A primary distinction among recovery methods is whether the anatomy is processed per volume or per cross-section. Volumetric models offer enhanced capture of global contextual information, but they can suffer from suboptimal learning due to elevated model complexity. Cross-sectional models with lower complexity offer improved learning behavior, yet they ignore contextual information across the longitudinal dimension of the volume. Here, we introduce a novel progressive volumetrization strategy for generative models (ProvoGAN) that serially decomposes complex volumetric image recovery tasks into successive cross-sectional mappings task-optimally ordered across individual rectilinear dimensions. ProvoGAN effectively captures global context and recovers fine-structural details across all dimensions, while maintaining low model complexity and improved learning behaviour. Comprehensive demonstrations on mainstream MRI reconstruction and synthesis tasks show that ProvoGAN yields superior performance to state-of-the-art volumetric and cross-sectional models.
Kokteyl partisi sırasında beyindeki doğal konuşma temsilleri
Humans are remarkably adept in selectively listening to a desired speaker in a crowded environment, while filtering out non-target speakers in the background. Attention is key to solving this difficult cocktail-party task, yet a detailed characterization of attentional effects on speech representations is lacking. It remains unclear across what levels of speech features and how much attentional modulation occurs in each brain area during the cocktail-party task. Besides, it should be clarified whether unattended speech is represented in cortex during selective listening and if so, at what feature levels its representations are maintained. To address these questions, we recorded whole-brain blood-oxygen-level-dependent (BOLD) responses while subjects either passively listened to single-speaker stories, or selectively attended to a male or a female speaker in temporally-overlaid stories in separate experiments. Spectral, articulatory, and semantic models of the natural stories were constructed to enable comprehensive assessments on the hierarchy of speech features. Intrinsic selectivity profiles were identified via voxelwise models fit to passive listening responses. Attentional modulations were then quantified based on model predictions for attended and unattended stories in the cocktail-party task. We find that acoustic representations are confined to the early auditory cortex whereas linguistic representations are broadly distributed across cortex, that attention causes broad modulations at multiple levels of speech representations (articulatory and semantic) while growing stronger towards later stages of processing, and that unattended speech is represented up to the semantic level in parabelt auditory cortex. These results provide insights on speech perception and attentional mechanisms that underlie the ability to selectively listen to a desired speaker in noisy multi-speaker environments.
Uzaysal destekli voksel-bazlı modelleme ve insan beyninde dinamik sahne kategorisi temsili
Humans have an impressive ability to rapidly process global information in natural scenes to infer their category. Yet, it remains unclear whether and how scene categories observed dynamically in the natural world are represented in cerebral cortex beyond few canonical scene-selective areas. To address this question, here we examined the representation of dynamic visual scenes by recording whole-brain blood oxygenation level-dependent (BOLD) responses while subjects viewed natural movies. We fit voxelwise encoding models to estimate tuning for scene categories that reflect statistical ensembles of objects and actions in the natural world. Voxelwise modeling (VM) is a powerful framework to predict single voxel responses evoked by a rich set of stimulus features present in complex natural stimuli. However, because VM disregards correlations across neighboring voxels, its sensitivity in detecting functional selectivity can be diminished in the presence of high levels of measurement noise. Here, we introduce spatially-informed voxelwise modeling (SPIN-VM) to take advantage of response correlations in spatial neighborhoods of voxels. To optimally utilize shared information, SPIN-VM performs regularization across spatial neighborhoods in addition to model features, while still generating single-voxel response predictions. Compared to VM, SPIN-VM yields higher prediction accuracies and better capture locally congruent information representations across cortex. We find that this scene-category model explains a significant portion of the response variance broadly across cerebral cortex. Cluster analysis of scene-category tuning profiles across cortex reveals nine spatially-segregated networks of brain regions consistently across subjects. These networks show heterogeneous tuning for a diverse set of dynamic scene categories related to navigation, human activity, social interaction, civilization, natural environment, non-human animals, motion-energy, and texture, suggesting that the organization of scene category representation is quite complex.
Derin denetimsiz öğrenme ile hızlandırılmış mrg rekonstrüksiyonu
Supervised reconstruction models are characteristically trained on matched pairs of undersampled and fully-sampled data to capture an MRI prior, along with supervision regarding the imaging operator to enforce data consistency. To reduce supervision requirements, the recent deep image prior framework instead conjoins untrained MRI priors with the imaging operator during inference. Yet, canonical convolutional architectures are suboptimal in capturing long-range relationships, and priors based on randomly initialized networks may yield suboptimal performance. To address these limitations, this thesis introduces a novel unsupervised MRI reconstruction method based on zero-Shot Learned Adversarial TransformERs (SLATER). SLATER embodies a deep adversarial network with cross-attention transformers to map noise and latent variables onto coil-combined MR images. During pre-training, this unconditional network learns a high-quality MRI prior in an unsupervised generative modeling task. During inference, a zero-shot reconstruction is then performed by incorporating the imaging operator and optimizing the prior to maximize consistency to undersampled data. Comprehensive experiments on brain MRI datasets clearly demonstrate the superior performance of SLATER against state-of-the-art unsupervised methods.
Derin üretken modeller ile federe MRG rekonstrüksiyonu
Multi-institutional efforts can facilitate training of deep MRI reconstruction models, albeit privacy risks arise during cross-site sharing of imaging data. Federated learning (FL) has recently been introduced to address privacy concerns by enabling distributed training without transfer of imaging data. Existing FL methods employ conditional reconstruction models to map from undersampled to fullysampled acquisitions via explicit knowledge of the accelerated imaging operator. Since conditional models generalize poorly across different acceleration rates or sampling densities, imaging operators must be fixed between training and testing, and they are typically matched across sites. To improve patient privacy, performance and flexibility in multi-site collaborations, here we introduce Federated learning of Generative IMage Priors (FedGIMP) for MRI reconstruction. FedGIMP leverages a two-stage approach: cross-site learning of a generative MRI prior, and prior adaptation following injection of the imaging operator. The global MRI prior is learned via an unconditional adversarial model that synthesizes high-quality MR images based on latent variables. A novel mapper subnetwork produces site-specific latents to maintain specificity in the prior. During inference, the prior is first combined with subject-specific imaging operators to enable reconstruction, and it is then adapted to individual cross-sections by minimizing a data-consistency loss. Comprehensive experiments on multi-institutional datasets clearly demonstrate enhanced performance of FedGIMP against both centralized and FL methods based on conditional models.
Kardiyak MRG rekonstrüksiyonu için bölütleme bilgisiyle desteklenen derin öğrenme algoritmaları
Deep learning methods have produced impressive results in accelerated magnetic resonance imaging (MRI) reconstruction from under-sampled k-space acquisitions. However, existing MRI reconstruction models are commonly trained with loss functions that uniformly weigh contributions from separate voxels across the field-of-view (FOV), without attributing focus on relatively important regions within the FOV. Furthermore common frameworks for model training rely on availability of large sets of fully-sampled MRI data to construct a ground-truth for the network output. This heavy reliance is undesirable as it is challenging to collect such large datasets in many applications, and even impossible for high spatiotemporal-resolution protocols. In this thesis, we first introduce a self-supervised learning methodology for dynamic cardiac MRI that trains the network to reconstruct acquisitions in the absence of fully-sampled data. We then introduce a segmentation-aware reconstruction framework which implicitly guides the reconstruction process around an ROI with the segmentation error signal. Lastly, we introduce RATNet, a reconstruction framework augmented with attention capabilities which explicitly carries spatial information into the reconstruction process to focus around regions of interest. Self-supervision reduces the excessive demand on fully-sampled data whereas the segmentation-aware reconstruction framework backpropagates the spatial information signal in to the reconstruction network. Lastly, RATNet incorporates the attention layers into reconstruction which are sensitive to focusing information supplied by the spatial information network. We demonstrate recovering fully-sampled images from under-sampled acquisitions in cardiac MRI and show their state-of-the-art performance in medical image reconstruction.
Manyetik parçacık görüntüleme için öğrenme tabanlı geriçatım teknikleri
Magnetic particle imaging (MPI) is a novel modality for imaging of magnetic nanoparticles (MNP) with high resolution, contrast and frame rate. An inverse problem is usually cased for reconstruction, which requires a time-consuming calibration scan for measuring a system matrix (SM). Previous calibration procedures involve scanning an MNP filled sample with a size that matches desired resolution through field of view. This time-consuming calibration scan which accounts for both system and MNP response imperfections is a critical factor prohibiting its practical use. Moreover, the quality of the reconstructed images heavily depend on the prior information about the MNP distribution as well as the specific reconstruction algorithm, since the inverse problem is highly ill-posed. Previous approaches commonly solve an optimization problem based on the measurement model that iteratively estimates the image while enforcing data consistency in an interleaved fashion. However, while conventional hand-crated priors do not fully capture the underlying complex features of MPI images, recently proposed learned priors suffer from limited generalization performance. To tackle these issues, we first propose a deep learning based technique for accelerated MPI calibration. The technique utilizes transformers for SM super-resolution (TranSMS) for accelerated calibration of SMs with high signal-to-noise-ratio. For signal-to-noise-ratio efficiency, we propose scanning a low resolution SM with larger MNP sample size. For improved SM estimation, TranSMS leverages the vision transformer to capture global contextual information while utilizing the convolutional module for local high-resolution features. Finally, a novel data-consistency module enforces measurement fidelity. TranSMS is shown to outperform competing methods significantly in terms of both SM recovery and image reconstruction performance. Next, to improve image reconstruction quality, we propose a novel physics-driven deep equilibrium based technique with learned consistency block for MPI (DEQ-MPI). DEQ-MPI embeds deep network operators into iterative optimization procedures for improved modeling of image statistics. Moreover, DEQ-MPI utilizes learned consistency to better capture the data statistics which helps improve the overall image reconstruction performance. Finally, compared to previous unrolling-based techniques, DEQ-MPI leverages implicit layers which enables training on the converged output. Demonstrations on both simulated and experimental data show that DEQ-MPI significantly improves image quality and reconstruction time over state-of-the-art reconstructions based on hand-crafted or learned priors.
Çok-kipli tıbbi görüntü sentezi için yeni derin öğrenme algoritmaları
Multi-modal medical imaging is a powerful tool for diagnosis and treatment of various diseases, as it provides complementary information about tissue morphology and function. However, acquiring multiple images from different modalities or contrasts is often impractical or impossible due to various factors such as scan time, cost, and patient comfort. Medical image translation has emerged as a promising solution to synthesize target-modality images given source-modality images. Ability to synthesize unavailable images enhance the ubiquity and utility of multi-modal protocols while decreasing examination costs and toxicity exposure such as ionizing radiation and contrast agents. Existing medical image translation methods prominently rely on generative adversarial networks (GANs) with convolutional neural networks (CNNs) backbones. CNNs are designed to perform local processing with compact filters, and this inductive bias is prone to limited contextual sensitivity. Meanwhile, GANs suffer from limited sample fidelity and diversity due to one-shot sampling and implicit characterization of the image distribution. To overcome the challenges with CNN based GAN models, in this thesis, first ResViT was introduced that leverages novel aggregated residual transformer (ART) blocks that synergistically fuse representations from convolutional and transformer modules. Then SynDiff is introduced, a conditional diffusion model that progressively maps noise and source images onto the target image via large diffusion steps and adversarial projections, capturing a direct correlate of the image distribution and improving sample quality and speed. ResViT provides a unified implementation to avoid the need to rebuild separate synthesis models for varying source-target modality configurations, whereas SynDiff enables unsupervised training on unpaired datasets via a cycle-consistent architecture. ResViT and SynDiff was demonstrated on synthesizing missing sequences in multi-contrast MRI, and CT images from MRI, and their state-of-the-art performance in medical image translation was shown.