Showing papers by "Paul Morris published in 2019"

Virtual Coronary Intervention: A Treatment Planning Tool Based Upon the Angiogram

Abstract: Objectives This study sought to assess the ability of a novel virtual coronary intervention (VCI) tool based on invasive angiography to predict the patient’s physiological response to stenting. Background Fractional flow reserve (FFR)-guided percutaneous coronary intervention (PCI) is associated with improved clinical and economic outcomes compared with angiographic guidance alone. Virtual (v)FFR can be calculated based upon a 3-dimensional (3D) reconstruction of the coronary anatomy from the angiogram, using computational fluid dynamics (CFD) modeling. This technology can be used to perform virtual stenting, with a predicted post-PCI FFR, and the prospect of optimized treatment planning. Methods Patients undergoing elective PCI had pressure-wire–based FFR measurements pre- and post-PCI. A 3D reconstruction of the diseased artery was generated from the angiogram and imported into the VIRTUheart workflow, without the need for any invasive physiological measurements. VCI was performed using a radius correction tool replicating the dimensions of the stent deployed during PCI. Virtual FFR (vFFR) was calculated pre- and post-VCI, using CFD analysis. vFFR pre- and post-VCI were compared with measured (m)FFR pre- and post-PCI, respectively. Results Fifty-four patients and 59 vessels underwent PCI. The mFFR and vFFR pre-PCI were 0.66 ± 0.14 and 0.68 ± 0.13, respectively. Pre-PCI vFFR deviated from mFFR by ±0.05 (mean Δ = −0.02; SD = 0.07). The mean mFFR and vFFR post-PCI/VCI were 0.90 ± 0.05 and 0.92 ± 0.05, respectively. Post-VCI vFFR deviated from post-PCI mFFR by ±0.02 (mean Δ = −0.01; SD = 0.03). Mean CFD processing time was 95 s per case. Conclusions The authors have developed a novel VCI tool, based upon the angiogram, that predicts the physiological response to stenting with a high degree of accuracy.

Evaluation of maxout activations in deep learning across several big data domains

Abstract: This study investigates the effectiveness of multiple maxout activation function variants on 18 datasets using Convolutional Neural Networks. A network with maxout activation has a higher number of trainable parameters compared to networks with traditional activation functions. However, it is not clear if the activation function itself or the increase in the number of trainable parameters is responsible in yielding the best performance for different entity recognition tasks. This paper investigates if an increase in the number of convolutional filters on traditional activation functions performs equal-to or better-than maxout networks. Our experiments compare the Rectified Linear Unit, Leaky Rectified Linear Unit, Scaled Exponential Linear Unit, and Hyperbolic Tangent activations to four maxout function variants. We observe that maxout networks train relatively slower than networks with traditional activation functions, e.g. Rectified Linear Unit. In addition, we found that on average, across all datasets, the Rectified Linear Unit activation function performs better than any maxout activation when the number of convolutional filters is increased. Furthermore, adding more filters enhances the classification accuracy of the Rectified Linear Unit networks, without adversely affecting their advantage over maxout activations with respect to network-training speed.

Maxout Neural Network for Big Data Medical Fraud Detection

Abstract: Globally, health care losses due to fraud rise every year, and for this reason fraud detection is an active research area that, in the U.S. alone, can potentially save billions of dollars. We explore the performance of multiple maxout activation variants on the big data medical fraud detection task using neural networks. Maxout networks have gained great success in many computer vision tasks, but there is limited work on other classification tasks. Our experiments compare Rectified Linear Unit, Leaky Rectified Linear Unit, Scaled Exponential Linear Unit, and hyperbolic tangent to four maxout variants. We evaluate the effectiveness of the activation functions on four U.S. Centers for Medicare and Medicaid Services datasets. Throughout this paper, we found that maxout networks are considerably slower to train compared to traditional activation functions. We find that on average, across all datasets, Scaled Exponential Linear Unit’s classification performance is better than any maxout activation, and reported the lowest training time.

Personalised fractional flow reserve: a novel concept to optimise myocardial revascularisation.

Abstract: AIMS Fractional flow reserve (FFR) represents the percentage reduction in coronary flow relative to a hypothetically normal artery; however, percutaneous coronary intervention (PCI) seldom achieves physiological normality (FFR 1.00), particularly in the context of diffuse disease. In this study we describe a method for calculating the vessel-specific maximal achievable FFR (FFRmax) providing a personalised assessment of what PCI can achieve. METHODS AND RESULTS FFR measurements were obtained from 71 patients (100 arteries) undergoing angiography. Three-dimensional (3D) coronary anatomy was reconstructed from angiographic images. An ideal intervention, in which all stenoses are removed, was modelled, and the FFRmax calculated. The "personalised" FFR (FFRpers) was calculated as measured FFR/FFRmax. PCI was performed in 52 vessels and post-PCI FFR measured in 50. FFRmax was compared to post-PCI measured FFRs. The mean FFRmax was 0.92 (±0.04). This was on average 0.04 (±0.05) higher than the corresponding post-PCI measured FFR (p<0.001). FFRpers was significantly higher (0.06±0.04) than measured FFR (p<0.001), indicating that FFR overestimates flow restoration achievable with PCI. CONCLUSIONS A patient's maximal achievable FFR can now be determined prior to PCI. This approach provides a more realistic assessment of the physiological benefit of PCI than is implied by baseline FFR and may prevent unnecessary intervention.

Characterization of Membrane Transporters by Heterologous Expression in E. coli and Production of Membrane Vesicles

Abstract: Several methods have been developed to functionally characterize novel membrane transporters. Polyamines are ubiquitous in all organisms, but polyamine exchangers in plants have not been identified. Here, we outline a method to characterize polyamine antiporters using membrane vesicles generated from the lysis of Escherichia coli cells heterologously expressing a plant antiporter. First, we heterologously expressed AtBAT1 in an E. coli strain deficient in polyamine and arginine exchange transporters. Vesicles were produced using a French press, purified by ultracentrifugation and utilized in a membrane filtration assay of labeled substrates to demonstrate the substrate specificity of the transporter. These assays demonstrated that AtBAT1 is a proton-mediated transporter of arginine, γ-aminobutyric acid (GABA), putrescine and spermidine. The mutant strain that was developed for the assay of AtBAT1 may be useful for the functional analysis of other families of plant and animal polyamine exchangers. We also hypothesize that this approach can be used to characterize many other types of antiporters, as long as these proteins can be expressed in the bacterial cell membrane. E. coli is a good system for the characterization of novel transporters, since there are multiple methods that can be employed to mutagenize native transporters.

Maxout Networks for Visual Recognition

Abstract: This study investigates the effectiveness of multiple maxout activation variants on image classification, facial identification and verification tasks using convolutional neural networks. A network with maxout activation has a higher number of trainable parameters compared to networks with traditional activation functions. However, it is not clear if the activation function itself or the increase in the number of trainable parameters is responsible for yielding the best performance on different entity recognition tasks. This article investigates if an increase in the number of convolutional filters on the rectified linear unit activation performs equal-to or better-than maxout networks. Our experiments compare rectified linear unit, leaky rectified linear unit, scaled exponential linear unit, and hyperbolic tangent to four maxout variants. Throughout the experiments, we found that on average, across all datasets, the rectified linear unit networks perform better than any maxout activation when the number of convolutional filters is increased six times.

17 Computational modelling of fractional flow reserve from coronary angiography : expert training required

Abstract: Lal K, Gosling R, Priest J, Stephenson T, Lee T, Robinson N, O’Connor T, Gregory B, Son S, Hodgson A, Dunnill J, Lawford P, Hose R, Morris PD, Gunn J Sheffield Teaching Hospitals NHS Foundation Trust and Department of Infection, Immunity and Cardiovascular Disease, University of Sheffield Introduction Visual estimation of the physiological significance of coronary artery disease (CAD) is inaccurate. Fractional flow reserve (FFR) is better but is under-used. A less invasive alternative is ‘virtual’ FFR (vFFR) calculated from computational fluid dynamics (CFD) modelling from angiographic images. The aim of this study was to quality assess the vFFRs analysed by non-expert operators by comparing their results to those of fully trained experts. Methods Two expert operators re-processed vFFRs from patients with CAD that had previously been processed by seven non-experts. The vFFRs were computed using the VIRTUheart™ tool (University of Sheffield). Figure 1 shows an example from the workflow. The vFFR results of the expert and non-expert analysed were compared on the basis of the recommendation for percutaneous coronary intervention vs medical therapy and the reason for the differences were documented. Inter- and intra-expert differences and the impact of the expert decisions upon potential clinical management were also assessed. Results The angiograms from 1098 patients with CAD were screened, from which 316 cases for vFFR analysis were identified as being suitable for processing. From these, one expert selected 264 consecutive cases for re-processing at random, of which 214 were successfully re-processed. Reasons for unsuccessful segmentation included inadequate images, poor opacification, overlap of vessels and unworkable geometry. The expert mean vFFR was 0.76 and the non-expert was 0.75 (mean per case difference 0.11, SD 0.12), with 73% agreement and 27% disagreement about treatment strategy (see figure 2). Of those, 18% would have been incorrectly revascularised and 9% incorrectly managed conservatively. The mean inter-observer (1st vs 2nd expert) and intra-observer (1st vs 1st expert) differences were 0.06 and 0.09 respectively, and agreement in management interpretations 89% and 90% respectively (p Conclusion There is a large difference in vFFR modelling between expert and less expert modellers. The differences are due to errors in 3-D vessel construction. There is little inter- or intra-observer variation between expert modellers. However good the modelling system, training is required to produce accurate vFFR results. Expert vFFR can improve the clinical management of patients with CAD, altering revascularisation decision in 37% cases. Conflict of Interest None

16 Assessing the accuracy of a novel in silico imaging tool for the 3D reconstruction of coronary vasculature in the context of virtual fractional flow reserve

Abstract: Fractional flow reserve (FFR) is the gold standard method for guiding percutaneous coronary intervention. ‘Virtual’ FFR (vFFR) offers a less-invasive alternative but accuracy is critically dependent on accurate 3D arterial reconstruction. This is especially challenging with angiography-based solutions due to practical challenges relating to image acquisition, notably table movement between image acquisitions. Some existing methods rely upon restricting table movement, but this poses difficulty in clinical practice. The aim of this study was to validate a novel method for 3D coronary arterial reconstruction under clinically realistic conditions. Six branched coronary arterial models (3 left and 3 right, 15 vessels) were generated in silico using patient angiograms and 3D printed in PLA (RepRap X400 PRO). All physical models underwent standard coronary angiography imaging. Each model was imaged three times with different restrictions on table movement (18 image datasets, 45 single-vessels). For 3D reconstruction, vessel centrelines were manually traced on two images >30° apart; automatic detection of the borders and diameter optimisation followed (Figure 1). All reconstructions were subjected to vFFR computation. Reconstructions were compared to the reference 3D files in terms of surface similarity (defined using Hausdorff measurements; averaged distance between a randomised sample of points on both meshes) and physiological analysis (vFFR). The effect of surface reconstruction error on physiological accuracy (vFFR) was described using Pearson’s correlation coefficient. To assess accuracy of diameter capture, three aluminium coronary phantoms were fabricated with concentric and eccentric stenoses (diameter range 0.74–1.77mm, % narrowing: 44.7–77.2%). These phantoms also underwent angiography and 3D reconstruction as previously described. Reconstructions were compared with physical micrometer measurements of percentage stenosis and minimum diameter. Accuracy was expressed as mean delta (±SD) and absolute error. Forty-five single-vessel reconstructions were analysed (Figure 2). The average distance between reconstructed and reference meshes (reconstruction error) was 0.65mm (±0.30) indicating excellent similarity throughout variation of table movement. Mean vFFR was 0.94 (±0.049) with an average absolute error of 0.008 ±0.0098 and a maximum absolute error of ±0.03. A weak positive relationship between error in reconstruction and physiology was demonstrated (r = 0.370, p = 0.013). Mean error of stenosis estimation using the metal phantoms was 1.2% (±1.2%). Accuracy of diameter reconstruction at maximum stenosis (minimum diameter) was excellent, with an error of 0.02mm (±0.06mm). Coronary anatomy can be reconstructed under realistic conditions with an accuracy that is acceptable for clinical decision-making. This novel method has the potential to facilitate interventional decision making as part of a vFFR workflow and may also have value in other areas of anatomical reconstruction. Conflict of Interest n/a

Deep Learning with Maxout Activations for Visual Recognition and Verification

Abstract: Visual recognition is one of the most active research topics in computer vision due to its potential applications in self-driving cars, healthcare, social media, manufacturing, etc. For image classification tasks, deep convolutional neural networks have achieved state-of-the-art results, and many activation functions have been proposed to enhance the classification performance of these networks. We explore the performance of multiple maxout activation variants on image classification, facial recognition and verification tasks using convolutional neural networks. Our experiments compare rectified linear unit, leaky rectified linear unit, scaled exponential linear unit, and hyperbolic tangent to four maxout variants. Throughout the experiments, we find that maxout networks train relatively slower than networks comprised of traditional activation functions. We found that on average, across all datasets, rectified linear units perform better than any maxout activation when the number of convolutional filters is increased six times.