Hemodynamic quantities are valuable biomedical risk factors for cardiovascular pathology such as atherosclerosis. Non-invasive, in-vivo measurement of these quantities can only be performed using a select number of modalities that are not widely available, such as 4D flow magnetic resonance imaging (MRI). In this work, we create a surrogate model for hemodynamic flow field estimation, powered by machine learning. We train graph neural networks that include priors about the underlying symmetries and physics, limiting the amount of data required for training. This allows us to train the model using moderately-sized, in-vivo 4D flow MRI datasets, instead of large in-silico datasets obtained by computational fluid dynamics (CFD), as is the current standard. We create an efficient, equivariant neural network by combining the popular PointNet++ architecture with group-steerable layers. To incorporate the physics-informed priors, we derive an efficient discretisation scheme for the involved differential operators. We perform extensive experiments in carotid arteries and show that our model can accurately estimate low-noise hemodynamic flow fields in the carotid artery. Moreover, we show how the learned relation between geometry and hemodynamic quantities transfers to 3D vascular models obtained using a different imaging modality than the training data. This shows that physics-informed graph neural networks can be trained using 4D flow MRI data to estimate blood flow in unseen carotid artery geometries.

Three-dimensional (3D) mesh reconstruction of the cardiac anatomy from medical images is useful for shape and motion measurements and biophysics simulations. However, 3D medical images are often acquired as 2D slices that are sparsely sampled (e.g., large slice spacing) and noisy, and 3D mesh reconstruction on such data is a challenging task. Traditional voxel-based approaches utilize non-differentiable pre- and post-processing that compromises fidelity to images, while mesh-level deep learning approaches require large 3D mesh annotations that are difficult to obtain. Differentiable cross-domain supervision from 2D images to 3D meshes is therefore crucial for enabling end-to-end optimization in medical imaging. While there have been attempts to approximate the voxelization and slicing of meshes that are being optimized, there has not yet been a method for directly using 2D slices to supervise 3D mesh reconstruction in a differentiable manner. Here, we propose a novel explicit differentiable voxelization and slicing (DVS) algorithm allowing gradient backpropagation to a 3D mesh from its slices, which facilitates refined mesh optimization directly supervised by the losses defined on 2D images. Further, we propose an innovative framework for extracting patient-specific left ventricle (LV) meshes from medical images by coupling DVS with a graph harmonic deformation (GHD) mesh morphing descriptor of cardiac shape that naturally preserves mesh quality and smoothness during optimization. The proposed framework achieves state-of-the-art performance in cardiac mesh reconstruction tasks from densely sampled (CT) as well as sparsely sampled (MRI stack with few slices) images, outperforming alternatives, including Marching Cubes, statistical shape models, algorithms with vertex-based mesh morphing algorithms and alternative methods for image-supervision of mesh reconstruction. Experimental results demonstrate that our method achieves an overall Dice score of 90% during a sparse fitting on multi-datasets. The proposed method can further quantify clinically useful parameters such as ejection fraction and global myocardial strains, closely matching the ground truth and outperforming the traditional voxel-based approach in sparse images.

For heart diagnostics, ultrasound is generally the modality of choice due to its high temporal and spatial resolution, availability, and patient safety. Although 3D echocardiography captures the complex shape and motion of the heart with more precision than 2D, it suffers to a greater extent from poor resolution, noise, and limited field-of-view. Multi-perspective echocardiography has proven to significantly enhance both image quality and field-of-view. The greatest improvements occur when combining acquisitions from widely differing insonification angles, but this process is challenging because of substantial local structural and brightness variations and ultrasound's anisotropic nature. To handle these inconsistencies, a novel temporal and spatial registration algorithm designed is proposed. Temporal registration is achieved using low-frequency cardiac wall features and motion extracted via singular value decomposition of a spatio-temporal Casorati matrix, while spatial registration is performed using phase-only correlation of low-frequency data. The acquisitions are seamlessly fused using a 3D, oriented, wavelet transform including a near-field clutter algorithm. In vitro and in vivo testing highlights the benefits of this approach. Temporal alignment, validated against electrocardiograms, is precise, with an average error of just 2 ± 10 ms. Furthermore, our method outperforms a six-degree-of-freedom encoder-based probe tracker, reducing spatial registration error to 5 ± 3 mm from 19 ± 10 mm. The resulting longitudinal and radial strain measurements closely align with those obtained by tagged magnetic resonance imaging, demonstrating the accuracy and feasibility of this technique.

In studies of Alzheimer's disease (AD), limited sample size considerably hampers the performance of intelligent diagnostic systems. Using multi-site data increases sample size but raises concerns regarding data privacy and inter-site heterogeneity. To address these issues, we developed a knowledge-guided federated graph attention learning network with a diffusion module to facilitate AD diagnosis from multi-site data. We used multiple templates to extract regions-of-interest (ROI)-based volume features from structural magnetic resonance imaging (sMRI) data. These volume features were then combined with previously identified AD features from published studies (prior knowledge) to determine the discriminative features within the images. We then designed an attention-guided diffusion module to synthesize samples by prioritizing these key features. The diffusion module was trained within a federated learning framework, which ensured inter-site data privacy while limiting data heterogeneity. Finally, we designed a federated graph attention learning network as a classifier to capture AD-related deep features and improve the accuracy of diagnosing AD. The efficacy of our approach was validated using three AD datasets. Thus, the classifier developed in this study represents a promising tool for optimizing multi-site neuroimaging data to improving the accuracy of diagnosing AD in the clinic.