Searcharxiv⌕ Search

arXiv subjects

Nicholas Hindley

Publications and source records attributed to Nicholas Hindley.

5 recordsLinked to original sources

Hyperelasticity as an efficient regularizer for 2D-3D thoracic motion models

Patient-specific motion models enable real-time volumetric imaging of the thorax from a single x-ray projection, but most existing approaches require hours of training. The VM-DS framework trains in minutes, yet its image-space self-consistency regularizer accounts for 40% of the motion-model fitting time and does not prevent the fitted fields from folding. In this paper, we explore a range of regularization approaches including scaling-and-squaring, bending energy, diffusion, first-order inverse consistency, isotropic total variation, a negative-Jacobian hinge, |log det J|, a control-point injectivity bound and hyperelastic regularization. Hyperparameters were selected on two development cases from the public DIR-Lab 4D dataset under rules fixed in advance, and the regularizers were tested on eight held-out cases. At a fixed number of training steps, the hyperelastic term trained in 5.6 against 8.4 min and gave a lower held-out TRE (2.5 mm). With a quarter of the motion-model budget it matched the baseline (2.7 mm), and at every budget from 2 to 17 min its mean TRE was lower by 0.3-0.5 mm. Hyperelastic fields had no folded voxel in any held-out view, whereas baseline fields folded in 83 % of views, with image similarity unchanged. A hyperelastic volume penalty can therefore replace image-space self-consistency, matching accuracy with 3-5 minutes of training and removing folding, bringing on-the-day motion models closer to routine use in image-guided interventions. Code will be available at https://github.com/Image-X-Institute/Voxelmap/tree/hyper

physics.med-ph↗

Just-in-time volumetric imaging: Learning 2D-3D thoracic motion models during image-guided interventions

Despite two decades of innovation, cutting-edge bronchoscopic techniques still fail to yield clear diagnostic decisions in 30% of cases and miss malignancies in 20% of cases. Intraprocedural 3D imaging drastically improves both tool positioning and diagnostic yield. However, each 3D spin represents a static snapshot of an anatomy that is continuously evolving as the patient is ventilated throughout the procedure. Here, we propose just-in-time volumetric imaging, in which a patient-specific 2D--3D motion model is trained on pre-intervention scans before it is needed and then supplies volumetric images from single fluoroscopic projections during sampling. Given the inspiration-hold CBCT already acquired for navigation and an additional motion-inclusive CBCT, our optimization regime named VoxelMap-DualScan (VM-DS), solves a low-rank motion model by alternating least squares and then amortizes it into a projection encoder trained on synthesized projections with known motion, all within a 10-minute window. On eight held-out cases from the public DIR-Lab 4DCT dataset, VM-DS achieved a mean landmark target registration error of 2.8 mm, a structural similarity index of 0.95 and a peak signal-to-noise ratio of 37.7 dB. Accuracy was maintained when reducing the number of projections used for training from 227 to 23. Inference takes 1.5 ms per projection, making the framework suitable for real-time use during image-guided interventions. Code will be made available at https://github.com/Image-X-Institute/Voxelmap/tree/VM-DS

physics.med-ph↗

Enabling Real-Time Volumetric Imaging in Interventional Radiology Suits via a Deep Learning Framework Robust to C-arm Tilt

Contemporary interventional imaging lacks the real-time 3D guidance needed for the precise localization of mobile thoracic targets. While Cone-Beam CT (CBCT) provides 3D data, it is often too slow for dynamic motion tracking. Deep learning frameworks that reconstruct 3D volumes from sparse 2D projections offer a promising solution, but their performance under the geometrically complex, non-zero tilt acquisitions common in interventional radiology is unknown. This study evaluates the robustness of a patient-specific deep learning framework, designed to estimate 3D motion, to a range of C-arm cranial-caudal tilts. Using a 4D digital phantom with a simulated respiratory cycle, 2D X-ray projections were simulated at five cranial-caudal tilt angles across 10 breathing phases. A separate deep learning model was trained for each tilt condition to reconstruct 3D volumetric images. The framework demonstrated consistently high-fidelity reconstruction across all tilts, with a mean Structural Similarity Index (SSIM) > 0.980. While statistical analysis revealed significant differences in performance between tilt groups (p < 0.0001), the absolute magnitude of these differences was minimal (e.g., the mean absolute difference in SSIM across all tilt conditions was ~0.0005), indicating they were not functionally significant. The magnitude of respiratory motion was found to be the dominant factor influencing accuracy, with the impact of C-arm tilt being a much smaller, secondary effect. These findings demonstrate that a patient-specific, motion-estimation-based deep learning approach is robust to geometric variations encountered in realistic clinical scenarios, representing a critical step towards enabling real-time 3D guidance in flexible interventional settings.

physics.med-ph↗

A statistical learning framework for mapping indirect measurements of ergodic systems to emergent properties

The discovery of novel experimental techniques often lags behind contemporary theoretical understanding. In particular, it can be difficult to establish appropriate measurement protocols without analytic descriptions of the underlying system-of-interest. Here we propose a statistical learning framework that avoids the need for such descriptions for ergodic systems. We validate this framework by using Monte Carlo simulation and deep neural networks to learn a mapping between low-field nuclear magnetic resonance spectra and proton exchange rates in ethanol-water mixtures. We found that trained networks exhibited normalized-root-mean-square errors of less than 1% for exchange rates under 150 s-1 but performed poorly for rates above this range. This differential performance occurred because low-field measurements are indistinguishable from one another at fast exchange. Nonetheless, where a discoverable relationship between indirect measurements and emergent dynamics exists, we demonstrate the possibility of approximating it without the need for precise analytic descriptions, allowing experimental science to flourish in the midst of ongoing theoretical work

physics.data-an↗

On Real-time Image Reconstruction with Neural Networks for MRI-guided Radiotherapy

MRI-guidance techniques that dynamically adapt radiation beams to follow tumor motion in real-time will lead to more accurate cancer treatments and reduced collateral healthy tissue damage. The gold-standard for reconstruction of undersampled MR data is compressed sensing (CS) which is computationally slow and limits the rate that images can be available for real-time adaptation. Here, we demonstrate the use of automated transform by manifold approximation (AUTOMAP), a generalized framework that maps raw MR signal to the target image domain, to rapidly reconstruct images from undersampled radial k-space data. The AUTOMAP neural network was trained to reconstruct images from a golden-angle radial acquisition, a benchmark for motion-sensitive imaging, on lung cancer patient data and generic images from ImageNet. Model training was subsequently augmented with motion-encoded k-space data derived from videos in the YouTube-8M dataset to encourage motion robust reconstruction. We find that AUTOMAP-reconstructed radial k-space has equivalent accuracy to CS but with much shorter processing times after initial fine-tuning on retrospectively acquired lung cancer patient data. Validation of motion-trained models with a virtual dynamic lung tumor phantom showed that the generalized motion properties learned from YouTube lead to improved target tracking accuracy. Our work shows that AUTOMAP can achieve real-time, accurate reconstruction of radial data. These findings imply that neural-network-based reconstruction is potentially superior to existing approaches for real-time image guidance applications.

physics.med-ph↗