Ali Bilgin

Leveraging diffusion tractography, connectivity-based parcellation (CBP) is one of the oldest methods for thalamic nuclei segmentation. The goal of this work was to reassess CBP using higher spatial resolution diffusion MRI data and reconstruction algorithms, and to compare it with recent state-of-the-art methods for thalamic nuclei segmentation. Furthermore, these methods were systematically evaluated against three histological atlases and one functional MRI–based atlas to examine their relative anatomical similarities and differences. High resolution diffusion and T1-weighted MRI data from 67 healthy individuals in the Human Connectome Project Young Adult database were analyzed. CBP was performed using probabilistic tractography with cortical targets derived from combining labels of the Human Connectome Project Multi-Modal Parcellation 1.0 atlas into 8, 11, and 23 regions. Results were compared against three recent methods: orientation distribution function clustering (ODF), track density imaging (TDI), and structural MRI-based segmentation. Group level analyses were conducted in the Montreal Neurological Institute space, and Dice overlap coefficients were calculated using four atlases (three histological, one functional). CBP results using newer data and methods were still remarkably similar to the original CBP parcellation results. Across atlases, a consistent hierarchy was observed: HIPS-THOMAS performed best, followed by TDI, ODF, and CBP (Kendall’s W = 1.00, p = 0.007). Histological atlases showed strong mutual agreement (Pearson r = 0.71–0.85), whereas the Zhang atlas demonstrated lower concordance (Pearson r = 0.51–0.63). Despite methodological advances, CBP remains constrained in its ability to delineate thalamic nuclei with histological accuracy. By contrast, structural and diffusion microstructural approaches provided better nuclear localization. These findings highlight the need for hybrid workflows that integrate structural and diffusion-based information to enable more reliable thalamic segmentation for neuroscience research.

The thalamus comprises multiple nuclei that support higher-order cognitive functions. However, its internal architecture remains difficult to delineate using conventional T1- or T2-weighted MRI because of limited tissue contrast. Diffusion-weighted MRI provides richer microstructural detail, yet accurate segmentation is still challenged by low anisotropy and tissue heterogeneity. To address these challenges, we present a modified spectral clustering framework for thalamic segmentation. Our approach jointly leverages voxel-wise information and fiber orientation distribution (FOD) features derived from multi-shell multi-tissue constrained spherical deconvolution. When evaluated using spatial probabilistic maps that capture across-subject spatial variability in labels, k-means and spectral clustering exhibit broadly similar group-level variability patterns. However, the spectral clustering framework accommodates smaller thalamic subdivisions, including the lateral and medial geniculate nuclei (LGN and MGN), which required exclusion from the k-means configuration for stable parcellation. Under these conditions, spectral clustering achieved Dice scores of 0.73 for the mediodorsal–parafascicular (MD–Pf) complex and 0.51 for the ventral posterolateral (VPL) nucleus and produce a cluster corresponding to LGN. Furthermore, by combining structural and diffusion information, our approach enabled subdivision of the pulvinar into four distinct regions. These result position our modified spectral clustering as a robust and anatomically informed tool for thalamic clustering and pulvinar sub-segmentation.

Purpose: To accelerate respiratory triggered free‐breathing T2 mapping of the abdomen while maintaining high‐quality anatomical images, accurate T2 maps, and fast reconstruction times.Methods: We developed a flexible deep learning framework that can be trained in a fully supervised manner to improve T2‐weighted images or in a self‐supervised manner to reconstruct T2 maps.Results: For retrospectively undersampled data, anatomical images and T2 maps reconstructed by the proposed deep learning method demonstrated reduced voxel‐wise error compared to existing traditional and compressed sensing techniques. Reconstruction times were approximately 1 s per slice, significantly faster than existing compressed sensing techniques. Prospectively undersampled data were also acquired to assess the model.Conclusion: The proposed deep‐learning framework reconstructed high‐quality anatomical images and accurate T2 maps from datasets undersampled to only 160 total radial views (5 views per echo time), enabling full liver coverage in under three minutes on average with per‐slice reconstruction times of approximately one second.

Diffusion tensor imaging (DTI) is essential for assessing brain microstructure but requires long acquisition times, limiting clinical use. Recent deep learning (DL) approaches, such as SuperDTI or deepDTI, improve DTI metrics but demand large, high-quality datasets for training. We propose a self-supervised deep learning with fine-tuning (SSDLFT) framework to reduce training data requirements. SSDLFT involves self-supervised pretraining, which denoises data without clean labels, followed by fine-tuning with limited high-quality data. Experiments using Human Connectome Project data show that SSDLFT outperforms traditional methods and other DL approaches in qualitative and quantitative assessments of DWI reconstructions and tensor metrics. SSDLFT’s ability to maintain high performance with fewer training subjects and DWIs presents a significant advancement, enhancing DTI’s practical applications in clinical and research settings.

Background: Contrastive learning, a successful form of representational learning, has shown promising results in pretraining deep learning (DL) models for downstream tasks. When working with limited annotation data, as in medical image segmentation tasks, learning domain‐specific local representations can further improve the performance of DL models.Purpose: In this work, we extend the contrastive learning framework to utilize domain‐specific contrast information from unlabeled Magnetic Resonance (MR) images to improve the performance of downstream MR image segmentation tasks in the presence of limited labeled data.Methods: The contrast in MR images is controlled by underlying tissue properties (e.g., T1 or T2) and image acquisition parameters. We hypothesize that learning to discriminate local representations based on underlying tissue properties should improve subsequent segmentation tasks on MR images. We propose a novel constrained contrastive learning (CCL) strategy that uses tissue‐specific information via a constraint map to define positive and negative local neighborhoods for contrastive learning, embedding this information in the representational space during pretraining. For a given MR contrast image, the proposed strategy uses local signal characteristics (constraint map) across a set of related multi‐contrast MR images as a surrogate for underlying tissue information. We demonstrate the utility of the approach for downstream: (1) multi‐organ segmentation tasks in T2‐weighted images where a DL model learns T2 information with constraint maps from a set of 2D multi‐echo T2‐weighted images (n = 101) and (2) tumor segmentation tasks in multi‐parametric images from the public brain tumor segmentation (BraTS) (n = 80) dataset where DL models learn T1 and T2 information from multi‐parametric BraTS images. Performance is evaluated on downstream multi‐label segmentation tasks with limited data in (1) T2‐weighted images of the abdomen from an in‐house Radial‐T2 (Train/Test = 30/20), (2) public Cartesian‐T2 (Train/Test = 6/12) dataset, and (3) multi‐parametric MR images from the public brain tumor segmentation dataset (BraTS) (Train/Test = 40/50). The performance of the proposed CCL strategy is compared to state‐of‐the‐art self‐supervised contrastive learning techniques. In each task, a model is also trained using all available labeled data for supervised baseline performance.Results: The proposed CCL strategy consistently yielded improved Dice scores, Precision, and Recall metrics, and reduced HD95 values across all segmentation tasks. We also observed performance comparable to the baseline with reduced annotation effort. The t‐SNE visualization of features for T2‐weighted images demonstrates its ability to embed T2 information in the representational space. On the BraTS dataset, we also observed that using an appropriate multi‐contrast space to learn T1+T2, T1, or T2 information during pretraining further improved the performance of tumor segmentation tasks.Conclusions: Learning to embed tissue‐specific information that controls MR image contrast with the proposed constrained contrastive learning improved the performance of DL models on subsequent segmentation tasks compared to conventional self‐supervised contrastive learning techniques. The use of such domain‐specific local representations could help understand, improve performance, and mitigate the scarcity of labeled data in MR image segmentation tasks.

Purpose: In radial abdominal imaging, it has been commonly observed that signal from the arms cause streaks due to system imperfections. We previously introduced a streak removal technique (B‐STAR), which is inherently spatially variant and limited to work in image space. In this work, we propose a spatially invariant streak cancellation technique (CACTUS), which can be applied in either image space or k‐space and is compatible with iterative reconstructions.Theory and Methods: Streak sources are typically spatially localized and can be represented using a low‐dimensional subspace. CACTUS identifies the streak subspace by leveraging the spatial redundancy of receiver coils and projects the data onto the streak null space to eliminate the streaks. When applied in k‐space, CACTUS can be combined with iterative reconstructions. CACTUS was tested in phantoms and in vivo abdominal imaging using a radial turbo spin‐echo pulse sequence.Results: In phantoms, CACTUS improved T2 estimation in comparison to previous de‐streaking methods. In vivo experiments showed that CACTUS reduced streaks and yielded T2 estimation, in regions affected by streaks, closer to a streak‐free reference. Evaluation using a clinical abdominal dataset (n = 20) showed that CACTUS is comparable to B‐STAR and yields significantly better signal preservation and streak cancellation than coil removal and suppression methods.Conclusion: CACTUS provides superior signal preservation and streak reduction performance compared to coil removal and suppression methods. As a clear advantage over B‐STAR, CACTUS can be integrated with iterative reconstruction methods. In abdominal T2 mapping, CACTUS improves the accuracy of parameter estimation in areas affected by streaks.

BACKGROUND AND PURPOSE: Accurate and reliable detection of white matter hyperintensities and their volume quantification can provide valuable clinical information to assess neurologic disease progression. In this work, a stacked generalization ensemble of orthogonal 3D convolutional neural networks, StackGen-Net, is explored for improving automated detection of white matter hyperintensities in 3D T2-FLAIR images. MATERIALS AND METHODS: Individual convolutional neural networks in StackGen-Net were trained on 2.5D patches from orthogonal reformatting of 3D-FLAIR (n = 21) to yield white matter hyperintensity posteriors. A meta convolutional neural network was trained to learn the functional mapping from orthogonal white matter hyperintensity posteriors to the final white matter hyperintensity prediction. The impact of training data and architecture choices on white matter hyperintensity segmentation performance was systematically evaluated on a test cohort (n = 9). The segmentation performance of StackGen-Net was compared with state-of-the-art convolutional neural network techniques on an independent test cohort from the Alzheimer’s Disease Neuroimaging Initiative-3 (n = 20). RESULTS: StackGen-Net outperformed individual convolutional neural networks in the ensemble and their combination using averaging or majority voting. In a comparison with state-of-the-art white matter hyperintensity segmentation techniques, StackGen-Net achieved a significantly higher Dice score (0.76 [SD, 0.08], F1-lesion (0.74 [SD, 0.13]), and area under precision-recall curve (0.84 [SD, 0.09]), and the lowest absolute volume difference (13.3% [SD, 9.1%]). StackGen-Net performance in Dice scores (median = 0.74) did not significantly differ (P = .22) from interobserver (median = 0.73) variability between 2 experienced neuroradiologists. We found no significant difference (P = .15) in white matter hyperintensity lesion volumes from StackGen-Net predictions and ground truth annotations. CONCLUSIONS: A stacked generalization of convolutional neural networks, utilizing multiplanar lesion information using 2.5D spatial context, greatly improved the segmentation performance of StackGen-Net compared with traditional ensemble techniques and some state-of-the-art deep learning models for 3D-FLAIR.

To develop and investigate a deep learning approach that uses sparse-view acquisition in dedicated breast computed tomography for radiation dose reduction, we propose a framework that combines 3D sparse-view cone-beam acquisition with a multi-slice residual dense network (MS-RDN) reconstruction. Projection datasets (300 views, full-scan) from 34 women were reconstructed using the FDK algorithm and served as reference. Sparse-view (100 views, full-scan) projection data were reconstructed using the FDK algorithm. The proposed MS-RDN uses the sparse-view and reference FDK reconstructions as input and label, respectively. Our MS-RDN evaluated with respect to fully sampled FDK reference yields superior performance, quantitatively and visually, compared to conventional compressed sensing methods and state-of-the-art deep learning based methods. The proposed deep learning driven framework can potentially enable low dose breast CT imaging.