Background and Objective : Micro computed tomography (microCT) enables detailed characterization of trabecular bone microarchitecture through quantitative morphological parameters. However, trabecular bone exhibits morphological variability across species, anatomical locations, bone types, and physiological or pathological conditions. Individual microCT datasets generally represent only a limited portion of this morphological variability, which may constrain the development and validation of quantitative descriptors, image-processing pipelines, and computational models. This study presents a Variational Autoencoder (VAE)-based framework for generating synthetic trabecular bone images, with the potential to support more comprehensive investigations of trabecular microarchitecture. Methods: The VAE is trained on individual microCT slices and learns a three-dimensional latent space in which the ordered sequence of slices is mapped onto a continuous trajectory. Generating a synthetic volume reduces to i) defining a latent space trajectory and ii) decoding its points into 2D images ensuring inter-slice coherence without requiring any explicit 3D architecture. An interactive trajectory editor allows the user to manipulate, extend, or resample the latent path, providing full control over the morphology, length, and resolution of the synthetic volume. The reconstruction capability of the VAE was assessed by comparing the original microCT slices with their VAE-based reconstructions, and the original microCT-derived volume with the corresponding 3D volume obtained from the VAE-based reconstructed images, using 2D and 3D descriptors of trabecular microarchitecture. New sequences of synthetic images could be generated by modifying latent-space trajectories through the editor, thereby enabling the generation of multiple synthetic 3D volumes with controlled morphological variability. In this study, one such sequence was generated, and the corresponding 3D synthetic volume was characterized using microarchitectural parameters. The resulting values were then compared with those reported in the literature to assess the plausibility of the generated structure. Results: Comparison of the morphological descriptors showed a high degree of similarity between the VAE-reconstructed and original microCT images. The corresponding 3D volumes also exhibited comparable morphological characteristics, indicating that the reconstruction process preserved the main structural features of the original microCT sample. Finally, the synthetic volume generated by synthetic VAE-based images exhibited microarchitectural parameter values comparable to those reported in the literature. Conclusion: These results provide support for the ability of the proposed framework to generate synthetic trabecular microarchitectures with coherent structural features. Therefore, the framework can support data augmentation for deep-learning model training and more extensive investigations of trabecular microarchitecture. Future developments will focus on linking selected morphological descriptors to the three latent-space axes, thereby expanding the range of microarchitectural variability that can be generated and investigated.
Background and objective Accurate quantification of muscle strain is essential for understanding musculoskeletal mechanics. Ultrasound-based methods for estimating muscle strain commonly rely on muscles with clearly visible fascicles and aponeuroses. We aimed to develop a novel ultrasound image analysis method for estimating muscle strain, particularly in muscles with indistinct aponeuroses. Methods Ultrasound B-mode images of the forearm flexor muscles were acquired during passive wrist extension in 16 healthy adults. Muscle strain was estimated using a feature-based tracking approach implemented using a Kanade–Lucas–Tomasi algorithm. Multiple feature points were automatically detected in the manually selected ellipse area at the initial frame, tracked across successive frames, and adaptively updated to accommodate time-varying tissue morphology. Displacement vectors were projected onto the first principal component of the tracked feature set; muscle strain was calculated from the spatially averaged longitudinal displacement field. Passive joint torque and shear-wave velocity were measured concurrently to examine the mechanical relevance of the estimated strain. Results Estimated muscle strain increased progressively with wrist extension and passive joint torque. At larger joint angles, participants with higher muscle stiffness exhibited smaller increases in strain, whereas those with lower stiffness demonstrated greater strain. Time-series analysis revealed significant group differences in strain responses during stretching. Conclusion The proposed B-mode image analysis method offers a preliminary framework for estimating muscle strain during passive stretching without relying on identifiable aponeuroses. Although the method successfully differentiates individual strain responses according to muscle stiffness, its validity and reliability require further enhancement before broad application.
Background and Objective Atrial Fibrillation (AF) is the most common cardiac arrhythmia and is associated with high risk of stroke and heart failure. While pulmonary vein isolation is a standard treatment, its efficacy is suboptimal for patients with persistent AF (PsAF), resulting in high recurrence rates. Increasing evidence indicates that fibrotic remodeling can create arrhythmogenic substrates outside the pulmonary veins, which may contribute to recurrence. Previous works have shown that PsAF Digital Twins (DT) can locate these arrhythmogenic regions and inform ablation targets for AF treatment, characterizing the full arrhythmogenic substrate through iterative simulations, including emergent reentries. While PsAF DTs are effective, their use in clinical practice is constrained by substantial computational demands. Methods In this study, we developed a Graph Neural Network (GNN) to directly identify regions sustaining reentrant drivers (ablation targets) without the need of computationally expensive simulations. The GNN was trained on 31 high-resolution PsAF DTs. Afterwards, the GNN's predictions were prospectively validated on four new PsAF DTs, where the predictions served as virtual ablation targets and new simulations were performed to evaluate the GNN’s predictions to eliminate reentrant drivers on new geometries and fibrotic distributions. Results Across 8-fold cross validation, the GNN achieved an AUROC of 0.884 and an AUPRC of 0.174, both of which surpassed all other tested models. GNN’s positive predictions with the greatest confidence were refined to regions with high fibrotic density and moderate fibrotic entropy. In the proof-of-concept prospective validation, GNN-guided ablation reduced emergent reentries by over 70% relative to the wide antral circumferential ablation pulmonary vein isolation (WACA PVI)-only control and reduced the mean number of additional simulation rounds required to reach non-inducibility by 66%. Conclusion This study demonstrates the potential of a GNN to guide DT-based ablation for PsAF patients and to reduce the iterative simulation burden of DT workflows. These findings support the feasibility of GNN-guided substrate prioritization, with further validation supporting clinical integration.
Background and objective High-frequency irreversible electroporation (H-FIRE) has shown promise in various medical applications. However, the complexity of the H-FIRE process makes it difficult to accurately predict the electric-field (E-field) lethal threshold, which is crucial for safe and effective treatment planning. Methods Here, we propose a prediction method integrated with deep learning (DL) and genetic algorithm (GA). Model interpretability is assessed using local interpretable model-agnostic explanation (LIME), and predictive performance is compared with traditional finite element simulation. Results The initial DL framework achieves a mean absolute error (MAE) of 0.02 on the training set and 0.03 on the test set, though some predictions deviate considerably from the true values. After GA optimization, the training MAE decreases to 0.013 (a 35% reduction) and the test MAE decreases to 0.023 (a 23% reduction). The maximal training error also shows a significant decrease in the lethal threshold from 1.4 to 0.3 kV·cm-1, with only 8% of test errors exceeding 0.2 kV·cm-1 (a 47% reduction). In addition, the optimized DL-GA model achieves a remarkable decrease in real MAE (0.06 kV·cm-1) compared to previous reports (0.511 kV·cm-1). Conclusion The proposed DL-GA framework enables rapid and accurate prediction of H-FIRE lethal thresholds, offering a potential tool for future clinical practice.
