Search

Data privacy is a major concern when accessing and processing sensitive medical data. A promising approach among privacy-preserving techniques is homomorphic encryption (HE), which allows for computations to be performed on encrypted data. Currently, HE still faces practical limitations related to high computational complexity, noise accumulation, and sole applicability the at bit or small integer values level. We propose herein an encoding method that enables typical HE schemes to operate on real-valued numbers of arbitrary precision and size. The approach is evaluated on two real-world scenarios relying on EEG signals: seizure detection and prediction of predisposition to alcoholism. A supervised machine learning-based approach is formulated, and training is performed using a direct (non-iterative) fitting method that requires a fixed and deterministic number of steps. Experiments on synthetic data of varying size and complexity are performed to determine the impact on runtime and error accumulation. The computational time for training the models increases but remains manageable, while the inference time remains in the order of milliseconds. The prediction performance of the models operating on encoded and encrypted data is comparable to that of standard models operating on plaintext data.

Background and Purpose— Therapeutic decision making for small unruptured intracranial aneurysms ( Methods— We analyzed 338 small unruptured aneurysms; 35 ruptured during the observation period, and 303 remained stable. Clinical, morphological, and hemodynamic parameters were considered. Computational fluid dynamics was used to calculate hemodynamic parameters based on computed tomography images of all aneurysms in their unruptured state. Differences between the ruptured and unruptured groups were tested by the Mann-Whitney U or Fisher exact tests. Multivariate logistic regression was applied to obtain a rupture risk model. Its predictive ability was investigated by receiver operating characteristic analysis. Results— The risk model revealed that rupture may be more likely to in younger patients (odds ratio [OR], 0.92 for each age increase of 1 year [95% CI, 0.88−0.96] P P =0.03), located at a bifurcation (OR, 5.45 [95% CI, 1.87−15.85] P =0.002), with a bleb (OR, 4.09 [95% CI, 1.42−11.79] P =0.009), larger length (OR, 1.91 for each increase of 1 mm [95% CI, 1.42−2.57] P P =0.01). The sensitivity, specificity, and area under the curve were 0.800, 0.752, and 0.826 (95% CI, 0.739−0.914) respectively. Conclusions— Younger age, presence of multiple aneurysms, location at a bifurcation, presence of a bleb, larger length, and lower pressure loss coefficient were identified as risk factors for rupture of small intracranial aneurysms. The risk model should be validated in further studies.

Funding Acknowledgements This research has been funded by the research grant PlaqueImage, contract number 26/01.09.2016, SMIS code 103544, Project funded by the European Union Aims Coronary shear stress (CSS) has been recently recognized to play a significant role in coronary plaque progression and vulnerabilisation. However, the evolution of CSS after implantation of different types of coronary stents is still under investigation. The aim of this study was to assess the evolution of the CSS along the coronary lesions following implantation of bioabsorbable vascular scaffolds (BVS), to determine the impact of BVS on coronary flow haemodynamics. Methods and results This was a single center prospective pilot study which enrolled 15 patients (aged 58.35 +/- 7.79 years, 13 males) with coronary artery disease undergoing BVS implantation in a major epicardial vessel. In all patients, angio CT scanning (Siemens Somatom Sensation scanner, Erlangen, Germany) was performed prior to the BVS implantation and repeated after 12 months. Lumen information was extracted from the vessels of interest and coronary rest hemodynamics, including CSS, were calculated using a computational fluid dynamics solver. All shear stress calculations were performed at baseline and repeated after 1 year. Average CSS was determined proximally, distally, and at the level of the minimal lumen area (MLA). Average CSS along the stented segment significantly decreased after BVS implantation, from 2.87 +/- 1.64 Pa at baseline to 1.9 +/- 0.49 at 1 year (p = 0.0001). Maximum CSS along the segment also exhibited a significant decrease, from 11.78 +/- 10.06 Pa to 6.35 +/- 3.08 Pa (p = 0.0009). Proximally to the MLA, CSS significantly decreased after BVS implantation, from 3.39 +/- 1.93 Pa at baseline to 1.91 +/- 0.68 Pa at 1 year (p Conclusions Implantation of BVS leads to a significant decrease of CSS after 1 year, especially within coronary segments located proximally to the stenosis. This underlines the role of BVS in re-establishing a physiological pattern of coronary flow in diseased coronary vessels. The feature (mentioned herein) is not commercially available. Due to regulatory reasons its future availability cannot be guaranteed.

In recent years, computational fluid dynamics (CFD) has become a valuable tool for investigating hemodynamics in cerebral aneurysms. CFD provides flow-related quantities, which have been shown to have a potential impact on aneurysm growth and risk of rupture. However, the adoption of CFD tools in clinical settings is currently limited by the high computational cost and the engineering expertise required for employing these tools, e.g., for mesh generation, appropriate choice of spatial and temporal resolution, and of boundary conditions. Herein, we address these challenges by introducing a practical and robust methodology, focusing on computational performance and minimizing user interaction through automated parameter selection. We propose a fully automated pipeline that covers the steps from a patient-specific anatomical model to results, based on a fast, graphics processing unit- (GPU-) accelerated CFD solver and a parameter selection methodology. We use a reduced order model to compute the initial estimates of the spatial and temporal resolutions and an iterative approach that further adjusts the resolution during the simulation without user interaction. The pipeline and the solver are validated based on previously published results, and by comparing the results obtained for 20 cerebral aneurysm cases with those generated by a state-of-the-art commercial solver (Ansys CFX, Canonsburg PA). The automatically selected spatial and temporal resolutions lead to results which closely agree with the state-of-the-art, with an average relative difference of only 2%. Due to the GPU-based parallelization, simulations are computationally efficient, with a median computation time of 40 minutes per simulation.

This chapter illustrates three clinical applications of the approaches presented so far in this book. The first one refers to a multi-scale, multi-physics cardiac modeling pipeline for cardiac resynchronization therapy (CRT), based solely on pre-operative, non-invasive data. Results on a small cohort of ten patients are presented and discussed. The second application introduces an AI based model for pressure drop computation in coarctation of the aorta (CoA). Since establishing a large enough patient-specific training database for AI model development would be prohibitively expensive and time consuming for this congenital pathology, a methodology for generating purely synthetic CoA anatomical models is described. The ground truth pressure drop values can be computed for the synthetic models using a three-dimensional Computational Fluid Dynamics (CFD) solver. The AI model reduces the mean absolute pressure drop prediction error more than five times compared to a previously published semi-analytical pressure drop model. Finally, the third application addresses the entire cardiovascular system, based on a lumped parameter model (LPM) of whole body circulation. A similar approach to the one employed for the CoA use case is considered, relying on a purely synthetic training database. Both time-independent, e.g. systemic resistance and compliance, and time-dependent quantities, e.g. ventricular pressure, are predicted. We show that the performance of the AI models is statistically similar to that of the LPM, while the inference time is reduced to milliseconds.

Motivated by state-of-the-art performances across a wide variety of areas, over the last few years Machine Learning has drawn a significant amount of attention from the healthcare domain. Despite their potential in enabling person-alized medicine applications, the adoption of Deep Learning based solutions in clinical workflows has been hindered in many cases by the strict regulations concerning the privacy of patient health data. We propose a solution that relies on Fully Homomorphic Encryption, particularly on the MORE scheme, as a mechanism for enabling computations on sensitive health data, without revealing the underlying data. The chosen variant of the encryption scheme allows for the computations in the Neural Network model to be directly performed on floating point numbers, while incurring a reasonably small computational overhead. For feasibility evaluation, we demonstrate on the MNIST digit recognition task that Deep Learning can be performed on encrypted data without compromising the accuracy. We then address a more complex task by training a model on encrypted data to estimate the outputs of a whole-body circulation (WBC) model. These results underline the potential of the proposed approach to outperform current solutions by delivering comparable results to the unencrypted Deep Learning based solutions, in a reasonable amount of time. Lastly, the security aspects of the encryption scheme are analyzed, and we show that, even though the chosen encryption scheme favors performance and utility at the cost of weaker security, it can still be used in certain practical applications.

Following the reports of breakthrough performances, machine learning based applications have become very popular in the medical field. However, with the recent increase in concerns related to data privacy, and the publication of specific regulations (e.g. GDPR), the development and, thus, exploitation of deep learning based applications in clinical decision making processes, has been rendered impossible in many cases. Herein, we describe and evaluate an approach that employs Fully Homo-morphic Encryption for allowing computations to be performed on sensitive data. Specifically, the solution exploits the MORE scheme and does not disclose patient data. The chosen encryption scheme increases the runtime only marginally and, importantly, allows for operations to be performed directly on floating-point numbers, which represents a critical property for artificial neural networks. The feasibility and performance are first evaluated on a standard benchmarking application (MNIST digit classification). Next, we considered a medical imaging application, i.e. classification of coronary views in X-ray angiography. The reported results indicate that the proposed solution has great potential: (i) computational results are indistinguishable from those obtained with the unencrypted variants of the deep learning-based applications, and (ii) run times increase only marginally. Finally, we also discuss in detail security concerns, and emphasize that the proposed solution may be employed in several practical applications, while still significant limitations remain to be solved in future work.

Owing to its clinical importance, there has been a growing body of research on understanding the hemodynamics of cerebral aneurysms. Traditionally, this work has been performed using general-purpose, state-of-the-art commercial solvers. This has meant requiring engineering expertise for making appropriate choices on the geometric discretization, time-step selection, choice of boundary conditions etc. Recently, a CFD research prototype has been developed (Siemens Healthcare GmbH, Prototype — not for diagnostic use) for end-to-end analysis of aneurysm hemodynamics. This prototype enables anatomical model preparation, hemodynamic computations, advanced visualizations and quantitative analysis capabilities. In this study, we investigate the accuracy of the hemodynamic solver in the prototype against a commercially available CFD solver ANSYS CFX 16.0 (ANSYS Inc., Canonsburg, PA, www.ansys.com) retrospectively on a sample of twenty patient-derived aneurysm models, and show good agreement of hemodynamic parameters of interest.

Solid voxelization represents the process of transforming a polygonal mesh into a voxel representation by associating each polygon of a mesh with the cells in the voxel grid. We introduce a novel approach for the voxelization of solid objects, designed for Graphics Processing Units (GPU). The method is based on a heuristic approach that computes an approximate distance field instead of using mesh surface normals or exact point-to-triangle distances. Two main steps are required: voxel marking and distance field computation. In the first step, each voxel is marked based on its location relative to the mesh (inside, outside of the domain or on its boundary), and, during the second step, a signed distance field is computed. Experiments focused on meshes encountered in medical imaging applications: a left ventricle and a coronary artery. The proposed method is found to be exceptionally robust as it is able to handle meshes with severe defects such as self intersections and holes. The GPU based implementation is on average 20 times faster than the multi-core CPU based implementation.

In this paper we introduce a methodology for performing one-way Fluid-Structure interaction (FSI), i.e. where the motion of the wall boundaries is imposed. We use a Graphics Processing Unit (GPU) accelerated Lattice-Boltzmann Method (LBM) implementation and present an efficient workflow for embedding the moving geometry, given as a set of polygonal meshes, in the LBM computation. The proposed method is first validated in a synthetic experiment: a vessel which is periodically expanding and contracting. Next, the evaluation focuses on the 3D Peristaltic flow problem: a fluid flows inside a flexible tube, where a periodic wave-like deformation produces a fluid motion along the centerline of the tube. Different geometry configurations are used and results are compared against previously published solutions. The efficient approach leads to an average execution time of approx. one hour per computation, whereas 50% of it is required for the geometry update operations. Finally, we also analyse the effect of changing the Reynolds number on the flow streamlines: the flow regime is significantly affected by the Reynolds number.

During the past decade Graphics Processing Units (GPU) have been increasingly employed for speeding up compute intensive scientific applications. In this field, the geometric multigrid method (GMG) is one of the most efficient algorithms for solving large sparse linear systems of equations. Herein we analyze the performance of an optimized GPU based implementation of the GMG method on different state-of-the-art NVIDIA GPUs. The GTX Titan Black card, set-up with increased double precision performance leads to the smallest execution time. It is marginally faster than the more recently released GTX Titan X card which has considerably lower double precision performance. Moreover, an energy efficiency analysis reveals that the GTX 660M and the more powerful Titan cards require a similar amount of energy for running the GMG algorithm: the larger execution time is compensated by the lower power consumption.