Diagnostic Performance of a Deep Learning Model Deployed at a National COVID-19 Screening Facility for Detection of Pneumonia on Frontal Chest Radiographs.

(1) Background: Chest radiographs are the mainstay of initial radiological investigation in this COVID-19 pandemic. A reliable and readily deployable artificial intelligence (AI) algorithm that detects pneumonia in COVID-19 suspects can be useful for screening or triage in a hospital setting. This study has a few objectives: first, to develop a model that accurately detects pneumonia in COVID-19 suspects; second, to assess its performance in a real-world clinical setting; and third, by integrating the model with the daily clinical workflow, to measure its impact on report turn-around time. (2) Methods: The model was developed from the NIH Chest-14 open-source dataset and fine-tuned using an internal dataset comprising more than 4000 CXRs acquired in our institution. Input from two senior radiologists provided the reference standard. The model was integrated into daily clinical workflow, prioritising abnormal CXRs for expedited reporting. Area under the receiver operating characteristic curve (AUC), F1 score, sensitivity, and specificity were calculated to characterise diagnostic performance. The average time taken by radiologists in reporting the CXRs was compared against the mean baseline time taken prior to implementation of the AI model. (3) Results: 9431 unique CXRs were included in the datasets, of which 1232 were ground truth-labelled positive for pneumonia. On the "live" dataset, the model achieved an AUC of 0.95 (95% confidence interval (CI): 0.92, 0.96) corresponding to a specificity of 97% (95% CI: 0.97, 0.98) and sensitivity of 79% (95% CI: 0.72, 0.84). No statistically significant degradation of diagnostic performance was encountered during clinical deployment, and report turn-around time was reduced by 22%. (4) Conclusion: In real-world clinical deployment, our model expedites reporting of pneumonia in COVID-19 suspects while preserving diagnostic performance without significant model drift.

Deep Supervised Domain Adaptation for Pneumonia Diagnosis From Chest X-Ray Images.

Pneumonia is one of the most common treatable causes of death, and early diagnosis allows for early intervention. Automated diagnosis of pneumonia can therefore improve outcomes. However, it is challenging to develop high-performance deep learning models due to the lack of well-annotated data for training. This paper proposes a novel method, called Deep Supervised Domain Adaptation (DSDA), to automatically diagnose pneumonia from chest X-ray images. Specifically, we propose to transfer the knowledge from a publicly available large-scale source dataset (ChestX-ray14) to a well-annotated but small-scale target dataset (the TTSH dataset). DSDA aligns the distributions of the source domain and the target domain according to the underlying semantics of the training samples. It includes two task-specific sub-networks for the source domain and the target domain, respectively. These two sub-networks share the feature extraction layers and are trained in an end-to-end manner. Unlike most existing domain adaptation approaches that perform the same tasks in the source domain and the target domain, we attempt to transfer the knowledge from a multi-label classification task in the source domain to a binary classification task in the target domain. To evaluate the effectiveness of our method, we compare it with several existing peer methods. The experimental results show that our method can achieve promising performance for automated pneumonia diagnosis.

Reference Ranges for Left Ventricular Curvedness and Curvedness-Based Functional Indices Using Cardiovascular Magnetic Resonance in Healthy Asian Subjects.

Curvature-based three-dimensional cardiovascular magnetic resonance (CMR) allows regional function characterization without an external spatial frame of reference. However, introduction of this modality into clinical practice is hampered by lack of reference values. We aim to establish normal ranges for 3D left ventricular (LV) regional parameters in relation to age and gender for 171 healthy subjects. LV geometrical reconstruction and automatic calculation of regional parameters were implemented by in-house software (CardioWerkz) using stacks of short-axis cine slices. Parameter normal ranges were stratified by gender and age categories (≤44, 45-64, 65-74 and 75-84 years). Our software had excellent intra- and inter-observer agreement. Ageing was significantly associated with increases in end-systolic (ES) curvedness (CES) and area strain (AS) with higher rates of increase in males, end-diastolic (ED) and ES wall thickness (WTED, WTES) with higher rates of increase in females, and reductions in ED and ES wall stress indices (σi,ED) with higher rates of increase in females. Females exhibited greater ED curvedness, CES, σi,ED and AS than males, but smaller WTED and WTES. Age × gender interaction was not observed for any parameter. This study establishes age and gender specific reference values for 3D LV regional parameters using CMR without additional image acquisition.

Multi-dimensional proprio-proximus machine learning for assessment of myocardial infarction.

This work presents a novel analysis methodology that utilises high-resolution, multi-dimensional information to better classify regions of the left ventricle after myocardial infarction. Specifically, the focus is to determine degree of infarction in regions of the left ventricle based on information extracted from cardiac magnetic resonance imaging. Enhanced classification accuracy is achieved using three mechanisms: Firstly, a plurality of indices/features is used in the pattern classification process, rather than a single index/feature (hence the term "multi-dimensional). Secondly, the method incorporates not only the indices/features of the region in consideration, but also indices/features from the neighbouring regions (hence the term "proprio-proximus"). Thirdly, advanced machine learning techniques are used for both feature selection and pattern classification process to ameliorate the effect of class-imbalance existing in the data. Numerical results from multiple experiments on real data showed that using multiple features improved the ability to distinguish between infarcted and non-infarcted remote segments, and using neighbouring information improved classification performance. The proposed methodology is general and can be adapted for the analysis of biological functions of other human organs.

Advanced analyses of computed tomography coronary angiography can help discriminate ischemic lesions.

BACKGROUND: Computed tomography coronary angiography (CTCA) image analysis enables plaque characterization and non-invasive fractional flow reserve (FFR) calculation. We analyzed various parameters derived from CTCA images and evaluated their associations with ischemia. METHODS: 49 (61 lesions) patients underwent CTCA and invasive FFR. Lesions with diameter stenosis (DS) ≥ 50% were considered obstructive. CTCA image processing incorporating analytical and numerical methods were used to quantify anatomical parameters of lesion length (LL) and minimum lumen area (MLA); plaque characteristic parameters of plaque volume, low attenuation plaque (LAP) volume, dense calcium volume (DCV), normalized plaque volume (NP Vol), plaque burden, eccentricity index and napkin-ring (NR) sign; and hemodynamic parameters of resistance index, stenosis flow reserve (SFR) and FFRB. Ischemia was defined as FFR ≤ 0.8. RESULTS: Plaque burden and plaque volume were inversely related to FFR. Multivariable logistic regression analysis identified the best anatomical, plaque and hemodynamic predictors, respectively, as DS (≥50% vs <50%; OR: 8.0; 95% CI: 1.6-39.4), normalized plaque volume (NP Vol) (≥4.3 vs <4.3; OR: 3.9; 95% CI: 1.1-14.0) and NR Sign (0 vs 1; OR: 13.6; 95% CI: 1.3-146.1), and FFRB (≤0.8 vs >0.8; OR: 44.4; 95% CI: 8.8-224.8). AUC increased from 0.70 with DS as the sole predictor to 0.81 after adding NP Vol and NR Sign; further addition of FFRB increased AUC to 0.93. CONCLUSION: Normalized plaque volume, napkin-ring derived from plaque analysis, and FFRB from numerical simulations on CTCA images substantially improved discrimination of ischemic lesions, compared to assessment by DS alone.

Human mesenchymal stem cell basal membrane bending on gratings is dependent on both grating width and curvature.

The topography of the extracellular substrate provides physical cues to elicit specific downstream biophysical and biochemical effects in cells. An example of such a topographical substrate is periodic gratings, where the dimensions of the periodic gratings influence cell morphology and directs cell differentiation. We first develop a novel sample preparation technique using Spurr's resin to allow for cross-sectional transmission electron microscopy imaging of cells on grating grooves, and observed that the plasma membrane on the basal surface of these cells can deform and bend into grooves between the gratings. We postulate that such membrane bending is an important first step in eliciting downstream effects. Thus, we use a combination of image analysis and mathematical modeling to explain the extent of bending of basal membrane into grooves. We show that the extent to which the basal membrane bends into grooves depends on both groove width and angle of the grating ridge. Our model predicts that the basal membrane will bend into grooves when they are wider than 1.9 µm in width. The existence of such a threshold may provide an explanation for how the width of periodic gratings may bring about cellular downstream effects, such as cell proliferation or differentiation.

Left Ventricular Wall Stress Is Sensitive Marker of Hypertrophic Cardiomyopathy With Preserved Ejection Fraction.

Hypertrophic cardiomyopathy (HCM) patients present altered myocardial mechanics due to the hypertrophied ventricular wall and are typically diagnosed by the increase in myocardium wall thickness. This study aimed to quantify regional left ventricular (LV) shape, wall stress and deformation from cardiac magnetic resonance (MR) images in HCM patients and controls, in order to establish superior measures to differentiate HCM from controls. A total of 19 HCM patients and 19 controls underwent cardiac MR scans. The acquired MR images were used to reconstruct 3D LV geometrical models and compute the regional parameters (i.e., wall thickness, curvedness, wall stress, area strain and ejection fraction) based on the standard 16 segment model using our in-house software. HCM patients were further classified into four quartiles based on wall thickness at end diastole (ED) to assess the impact of wall thickness on these regional parameters. There was a significant difference between the HCM patients and controls for all regional parameters (P < 0.001). Wall thickness was greater in HCM patients at the end-diastolic and end-systolic phases, and thickness was most pronounced in segments at the septal regions. A multivariate stepwise selection algorithm identified wall stress index at ED (σ i,ED ) as the single best independent predictor of HCM (AUC = 0.947). At the cutoff value σ i,ED < 1.64, both sensitivity and specificity were 94.7%. This suggests that the end-diastolic wall stress index incorporating regional wall curvature-an index based on mechanical principle-is a sensitive biomarker for HCM diagnosis with potential utility in diagnostic and therapeutic assessment.

Robust off-line heartbeat detection using ECG and pressure-signals.

Artefacts in pressure- and ECG-signals generally arise due to different causes. Therefore, the combined analysis of both signals can increase the effectiveness of heartbeat detection compared to analysis using solely ECG-signals. In this paper, we present an algorithm for heartbeat annotation by combining the analysis of both the pressure- and ECG-signals. The novelties of our algorithm are as follows: (1) development of a new approach for annotating heartbeats using pressure-signals, (2) development of a mechanism that identifies and corrects paced rhythms, and (3) development of a noise detection approach. Our algorithm is tested on the datasets from the extended phase of the Physionet CINC-2014 challenge and produces an overall score of 87.31%. Finally, we put forth several recommendations that could further improve our algorithm.

Volume Preserved Mass-Spring Model with Novel Constraints for Soft Tissue Deformation.

An interactive surgical simulation system needs to meet three main requirements, speed, accuracy, and stability. In this paper, we present a stable and accurate method for animating mass-spring systems in real time. An integration scheme derived from explicit integration is used to obtain interactive realistic animation for a multiobject environment. We explore a predictor-corrector approach by correcting the estimation of the explicit integration in a poststep process. We introduce novel constraints on positions into the mass-spring model (MSM) to model the nonlinearity and preserve volume for the realistic simulation of the incompressibility. We verify the proposed MSM by comparing its deformations with the reference deformations of the nonlinear finite-element method. Moreover, experiments on porcine organs are designed for the evaluation of the multiobject deformation. Using a pair of freshly harvested porcine liver and gallbladder, the real organ deformations are acquired by computed tomography and used as the reference ground truth. Compared to the porcine model, our model achieves a 1.502 mm mean absolute error measured at landmark locations for cases with small deformation (the largest deformation is 49.109 mm) and a 3.639 mm mean absolute error for cases with large deformation (the largest deformation is 83.137 mm). The changes of volume for the two deformations are limited to 0.030% and 0.057%, respectively. Finally, an implementation in a virtual reality environment for laparoscopic cholecystectomy demonstrates that our model is capable to simulate large deformation and preserve volume in real-time calculations.

Regional ejection fraction and regional area strain for left ventricular function assessment in male patients after first-time myocardial infarction.

In this work, we present a method to assess left ventricle (LV) regional function from cardiac magnetic resonance (CMR) imaging based on the regional ejection fraction (REF) and regional area strain (RAS). CMR scans were performed for 30 patients after first-time myocardial infarction (MI) and nine age- and sex-matched healthy volunteers. The CMR images were processed to reconstruct three-dimensional LV geometry, and the REF and RAS in a 16-segment model were computed using our proposed methodology. The method of computing the REF was tested and shown to be robust against variation in user input. Furthermore, analysis of data was feasible in all patients and healthy volunteers without any exclusions. The REF correlated well with the RAS in a nonlinear manner (quadratic fit-R(2) = 0.88). In patients after first-time MI, the REF and RAS were significantly reduced across all 16 segments (REF: p < 0.05; RAS: p < 0.01). Moreover, the REF and RAS significantly decreased with the extent of transmural scar obtained from late gadolinium-enhanced CMR images. In addition, we show that the REF and RAS can be used to identify regions with compromised function in the patients with preserved global ejection fraction with reasonable accuracy (more than 78%). These preliminary results confirmed the validity of our approach for accurate analysis of LV regional function. Our approach potentially offers physicians new insights into the local characteristics of the myocardial mechanics after a MI.

A virtual surgical training system that simulates cutting of soft tissue using a modified pre-computed elastic model.

This work presents a surgical training system that incorporates cutting operation of soft tissue simulated based on a modified pre-computed linear elastic model in the Simulation Open Framework Architecture (SOFA) environment. A precomputed linear elastic model used for the simulation of soft tissue deformation involves computing the compliance matrix a priori based on the topological information of the mesh. While this process may require a few minutes to several hours, based on the number of vertices in the mesh, it needs only to be computed once and allows real-time computation of the subsequent soft tissue deformation. However, as the compliance matrix is based on the initial topology of the mesh, it does not allow any topological changes during simulation, such as cutting or tearing of the mesh. This work proposes a way to modify the pre-computed data by correcting the topological connectivity in the compliance matrix, without re-computing the compliance matrix which is computationally expensive.

Design and construction of an equibiaxial cell stretching system that is improved for biochemical analysis.

We describe the design and validation of an equibiaxial stretching device in which cells are confined to regions of homogeneous strain. Using this device, we seek to overcome a significant limitation of existing equibiaxial stretching devices, in which strains are not homogeneous over the entire region of cell culture. We cast PDMS in a mold to produce a membrane with a cylindrical wall incorporated in the center, which was used to confine cell monolayers to the central membrane region subjected to homogeneous equibiaxial strain. We demonstrated that the presence of the wall to hold the culture medium did not affect strain homogeneity over the majority of the culture surface and also showed that cells adhered well onto the PDMS membranes. We used our device in cyclic strain experiments and demonstrated strain-dependent changes in extracellular signal-regulated kinase (ERK) and tyrosine phosphorylation upon cell stretching. Furthermore, we examined cell responses to very small magnitudes of strain ranging from 1% to 6% and were able to observe a graduated increase in ERK phosphorylation in response to these strains. Collectively, we were able to study cellular biochemical response with a high degree of accuracy and sensitivity to fine changes in substrate strain. Because we have designed our device along the lines of existing equibiaxial stretching technologies, we believe that our innovations can be incorporated into existing systems. This device would provide a useful addition to the set of tools applied for in vitro studies of cell mechanobiology.

Graph-cuts based reconstructing patient specific right ventricle: first human study.

Right ventricular (RV) function is increasingly recognized to play an important role in the clinical status and long-term outcome in patients with congenital heart disease as well as ischemic cardiomyopathy with left ventricular dysfunction. However, quantification of RV characteristics and function are still challenging due to its complex morphology and its thin wall with coarse trabeculations. To assess RV functions quantitatively, establishing the patient-specific model from medical images is a prerequisite task. This study aims to develop a novel method for RV model reconstruction. Magnetic resonance images were acquired and preprocessed. Contours of right ventricle, right atrium and pulmonary artery were manually delineated at all slices and all time frames. The contour coordinates as well as the medical image specifications such as image pixel resolution and slick thickness were exported. The contours were transformed to the correct positions. Reorientation and matching were executed in between neighboring contours; extrapolation and interpolation were conducted upon all contours. After preprocessing, the more dense point set was reconstructed through a variational tool. A Delaunay-based tetrahedral mesh was generated on the region of interest. The weighted minimal surface model was used to describe RV surface. The graphcuts technique, i.e., max-flow/min-cut algorithm, was applied to minimize the energy defined by the model. The reconstructed surface was extracted from the mesh according to the mincut. Smoothing and remeshing were performed. The CPU time to reconstruct the model for one frame was approximately 2 minutes. In 10 consecutive subjects referred for cardiac MRI (80% female), right ventricular volumes were measured using our method against the commercial available CMRtools package. The results demonstrated that there was a significant correlation in end-diastolic and end-systolic volumes between our method and commercial software (r= 0.89 for end-diastolic volume and r=0.79 for end-systolic volume, both P<;0.0001). The time to obtain right ventricular volumes was shorter using our method than commercial one. In conclusion, a new method for right ventricle reconstruction has been developed. We envisage that this automatic modeling tool could be used by radiographer and cardiologists to assess the RV function in diverse heart diseases.

Anisotropic rigidity sensing on grating topography directs human mesenchymal stem cell elongation.

Through mechanotransduction, cells can sense physical cues from the extracellular environment and convert them into internal signals that affect various cellular functions. For example, human mesenchymal stem cells (hMSCs) cultured on topographical gratings have been shown to elongate and differentiate to different extents depending on grating width. Using a combination of experiments and mathematical modeling, the physical parameters of substrate topography that direct cell elongation were determined. On a variety of topographical gratings with different grating widths, heights and rigidity, elongation of hMSCs was measured and a monotonic increase was observed for grating aspect ratio (crosssectional height to line-width ratio) between 0.035 and 2. The elongation was also dependent on the grating substrate rigidity over a range of 0.18-1.43 MPa. A mathematical model was developed to explain our observations by relating cell elongation to the anisotropic deformation of the gratings and how this anisotropy depends on the aspect ratio and rigidity of the gratings. Our model was in good agreement with the experimental data for the range of grating aspect ratio and substrate rigidity studied. In addition, we also showed that the percentage of aligned cells, which had a strong linear correlation with elongation for slightly elongated cells, saturated toward 100 % at higher level of cell elongation. Our results may be useful in designing gratings to elicit specific cellular responses that may depend on the extent of cell elongation.

Synchronous simulation for deformation of liver and gallbladder with stretch and compression compensation.

One challenge in surgical simulation is to design stable deformable models to simulate the dynamics of organs synchronously. In this paper, we develop a novel mass-spring model on the tetrahedral meshes for soft organs such as the liver and gallbladder, which can stably deform with large time steps. We model the contact forces between the organs as a kind of forces generated by the tensions of repulsive springs connecting in between the organs. The simulation system couples a pair of constraints on the length of springs with an implicit integration method. Based on the novel constraints, our simulator can efficiently preserve the volumes and geometric properties of the liver and gallbladder during the simulation. The numerical examples demonstrate that the proposed simulation system can provide realistic and stable deformable results.

A three-dimensional random network model of the cytoskeleton and its role in mechanotransduction and nucleus deformation.

We have developed a three-dimensional random network model of the intracellular actin cytoskeleton and have used it to study the role of the cytoskeleton in mechanotransduction and nucleus deformation. We use the model to predict the deformation of the nucleus when mechanical stresses applied on the plasma membrane are propagated through the random cytoskeletal network to the nucleus membrane. We found that our results agree with previous experiments utilizing micropipette pulling. Therefore, we propose that stress propagation through the random cytoskeletal network can be a mechanism to effect nucleus deformation, without invoking any biochemical signaling activity. Using our model, we also predict how nucleus strain and its relative displacement within the cytosol vary with varying concentrations of actin filaments and actin-binding proteins. We find that nucleus strain varies in a sigmoidal manner with actin filament concentration, while there exists an optimal concentration of actin-binding proteins that maximize nucleus displacement. We provide a theoretical analysis for these nonlinearities in terms of the connectivity of the random cytoskeletal network. Finally, we discuss laser ablation experiments that can be performed to validate these results in order to advance our understanding of the role of the cytoskeleton in mechanotransduction.

Cellular deformation and intracellular stress propagation during optical stretching.

Experiments have shown that mechanical stress can regulate many cellular processes. However, in most cases, the exact regulatory mechanisms are still not well understood. One approach in improving our understanding of such mechanically induced regulation is the quantitative study of cell deformation under an externally applied stress. In this paper, an axisymmetric finite-element model is developed and used to study the deformation of single, suspended fibroblasts in an optical stretcher in which a stretching force is applied onto the surface of the cell. A feature of our physical model is a viscoelastic material equation whose parameters vary spatially to mimic the experimentally observed spatial heterogeneity of cellular material properties. Our model suggests that cell size is a more important factor in determining the maximal strain of the optically stretched fibroblasts compared to the thickness of the actin cortical region. This result could explain the higher deformability observed experimentally for malignant fibroblasts in the optical stretcher. Our model also shows that maximal stress propagates into the nuclear region for malignant fibroblasts whereas for normal fibroblasts, it does not. We discuss how this may impact the transduction of cancer signaling pathways.