Search

This study evaluates the performance of convolutional neural networks (CNNs) in risk stratifying the malignant potential of thyroid nodules alongside traditional methods such as American College of Radiology Thyroid Imaging Reporting and Data System (ACR TIRADS). The data set consisted of 651 pathology-proven thyroid nodules (500 benign, 151 malignant) from 571 patients collected at a single tertiary academic medical center. Each thyroid nodule consisted of two orthogonal views (sagittal and transverse) for a total of 1,302 grayscale images. A CNN classifier was developed to identify malignancy versus benign thyroid nodules, and a nested double cross validation scheme was applied to allow for both model parameter selection and for model accuracy evaluation. All thyroid nodules were classified according to ACR TIRADS criteria and were compared with their respective CNN-generated malignancy scores. The best performing model was the MobileNet CNN ensemble with an area under the curve of 0.86 (95% confidence interval, 0.83-0.90). Thyroid nodules within the highest and lowest CNN risk strata had malignancy rates of 81.4% and 5.9%, respectively. The rate of malignancy for ACR TIRADS ranged from 0% for TR1 nodules to 60% for TR5 nodules. Convolutional neural network malignancy scores correlated well with TIRADS levels, as malignancy scores ranged from 0.194 for TR1 nodules and 0.519 for TR5 nodules. Convolutional neural networks can be trained to generate accurate malignancy risk scores for thyroid nodules. These predictive models can aid in risk stratifying thyroid nodules alongside traditional professional guidelines such as TIRADS and can function as an adjunct tool for the radiologist when identifying those patients requiring further histopathologic workup.

White matter hyperintensity (WMH) is an abnormal T2 signal in the deep and subcortical white matter visualized on MRI associated with hypertension, cerebrovascular disease, and aging. The Fazekas (Fz) scoring system is a commonly used qualitative tool to assess the severity of WMH. While studies have compared Fazekas scores to other scoring methods, the comparison of Fazekas scores and volume of WMH using current semiautomated volumetric techniques has not been studied.We reviewed MRI studies acquired at our institution between 2015 and 2017. Relative WMH was scored by one author trained in Fazekas scoring. A board certified neuroradiologist scored them independently for confirmation. Manual segmentations of WMH were completed using 3D Slicer 4.9. A 3D model was formed to quantify WMH in milliliters (mL). ANOVA tests were performed to determine the association of Fazekas scores with corresponding WMH volumes.Among the 198 patients in our study, WMH were visualized in 163 (Fz1: n=66; Fz2: n=49; Fz3: n=48). WMH volumes significantly differed according to Fazekas score (F = 141.1, p0.001), with increasing WMHV associated with higher Fazekas scores: Fz1, range 0.1-8.3 mL (mean 3.7, SD 2.3); Fz2, range 6.0-17.7 mL (mean 10.8, SD 3.1); Fz3, range 14.2-77.2 mL (mean 35.2, SD 17.9); and Fz3 (excluding 11 outliers above 50 mL), 14.2-47.0 mL (mean 27.1, SD 8.9).Fazekas scores correspond with distinct ranges of WMH volume with relatively little overlap, but scores based on volumes are more efficacious. A modified Fazekas from 0-4 should be considered.

We present a case study evaluating the potential for interactively identifying placental surface blood vessels using magnetic resonance imaging (MRI) scans in virtual reality (VR) environments. We visualized the MRI data using direct volume rendering in a high-fidelity CAVE-like VR system, allowing medical professionals to identify relevant placental vessels directly from volume visualizations in the VR system, without prior vessel segmentation. Participants were able to trace most of the observable vascular structure, and consistently identified blood vessels down to diameters of 1 mm, an important requirement in diagnosing vascular diseases. Qualitative feedback from our participants suggests that our VR visualization is easy to understand and allows intuitive data exploration, but complex user interactions remained a challenge. Using these observations, we discuss implications and requirements for spatial tracing user interaction methods in VR environments. We believe that VR MRI visualizations are the next step towards effective surgery planning for prenatal diseases.

Healthcare providers rely on complex biomedical devices to assess, treat, and monitor patients. Ongoing research efforts are attempting to generate and implement better algorithms and mechanisms to ensure the early, accurate, automated, and clinically meaningful recognition of patterns and changes in patient health and pathology. Effecting such evolutionary advances in patient monitoring will likely require large collections of high-resolution physiologic parameter datasets from a broad spectrum of patients. As part of a research program to scientifically improve patient monitoring (with a focus on alarm fatigue mitigation), investigators developed the Medical Technology Interface-Open/Research toolkit with modular conduit components that provide the following capabilities: 1) access to select bedside monitor physiologic signals in real-world clinical settings for near-real-time acquisition, storage, and export of high-resolution patient datastreams in a portable format (.json); 2) establishment of a safe, parallel test environment at the bedside for experimental datastream analyses in a research framework. Deployment and interfacing of toolkit elements with off-the-shelf software solutions in a live emergency department setting enabled the construction of a bedside clinical informatics (BCI) research pipeline infrastructure that featured 1) indexing, search/query, and retrieval of datastreams for sophisticated analyses, experimental processing, and algorithm development; and 2) dataset visualization for expert adjudication of datastream interpretability, alarm clinical significance and severity, and experimental algorithm performance. In order to help institute a collaborative biomedical engineering research resource, this article shares details of the active ED BCI data pipeline and presents preliminary examples of ongoing multimodal data fusion applications.

Automated segmentation of medical imaging is of broad interest to clinicians and machine learning researchers alike. The goal of segmentation is to increase efficiency and simplicity of visualization and quantification of regions of interest within a medical image. Image segmentation is a difficult task because of multiparametric heterogeneity within the images, an obstacle that has proven especially challenging in efforts to automate the segmentation of brain lesions from non-contrast head computed tomography (CT). In this research, we have experimented with multiple available deep learning architectures to segment different phenotypes of hemorrhagic lesions found after moderate to severe traumatic brain injury (TBI). These include: intraparenchymal hemorrhage (IPH), subdural hematoma (SDH), epidural hematoma (EDH), and traumatic contusions. We were able to achieve an optimal Dice Coefficient1 score of 0.94 using UNet++ 2D Architecture with Focal Tversky Loss Function, an increase from 0.85 using UNet 2D with Binary Cross-Entropy Loss Function in intraparenchymal hemorrhage (IPH) cases. Furthermore, using the same setting, we were able to achieve the Dice Coefficient score of 0.90 and 0.86 in cases of Extra-Axial bleeds and Traumatic contusions, respectively., Comment: 9 pages, 3 figures, 3 tables

Computed tomography (CT) is commonly used to assess traumatic brain injury (TBI) in the emergency department (ED). Radiologists at a Level 1 trauma center implemented a novel tool, the RADiology CATegorization (RADCAT) system, to communicate injuries to clinicians in real time. Using this categorization system, we aimed to determine the rates of positive head CTs among pediatric and adult ED patients evaluated for TBI.We performed a retrospective analysis of all patients who received a head CT to assess for TBI. We classified head CTs using the RADCAT tool. On a 5-point scale, scores of 3 or less are considered normal or routine. Scores of 4-5 are considered high priority, representing findings such as intracranial bleeding.Of the 5,341 head CT's obtained during the study period, 992 (18.5%) had high priority results (scores 4-5). A large number of pediatric studies, 30.8%, were positive for high priority results. Among the adult population, 18.0 % contained high priority results.The pediatric population had a higher rate of high priority reads among those undergoing non- contrast head CT for TBI compared to adult patients.

BACKGROUND Invasion of glioblastoma tumor cells beyond the visible margins is hypothesized to mediate tumor spread and recurrence, and thereby affect poor survival. Radiomic biomarkers associated with the extent of tumor infiltration are virtually non-existent. METHODS Seven subjects diagnosed with glioblastoma underwent high resolution “Q-Space” diffusion-weighted MRI sequences (Siemens 3T scanner, 64 gradient directions, b-value=1000s/mm2) during pre-operative MRI workup. For each subject, the largest tumor was manually segmented, and patterns of probabilistic inter-voxel coherence intersecting each segmentation generated as tractograms using DSI Studio (length=0–200mm, AT=30º). Separately, immunodeficient (Nu/J) mice were injected sub-hippocampally with patient-derived glioma stem cells (GSCs) using stereotactic guidance. Two months after injection, mice were sacrificed, resected brains were scanned on a Bruker 7T MRI using cryoprobe with T2-weighted and diffusion-weighted sequences (512 directions), and tumor-intersecting tractography displayed. 3D whole-brain reconstruction of the same xenograft model, stained with anti-human mitochondrial antibody, was performed for comparison. In one patient undergoing resection of fronto-temporal glioblastoma, BrainLab intraoperative navigation was used for stereotactic biopsy of extra-tumor parenchymal samples localized according to proximity to tumor-intersecting tractography. RNA-seq was performed on all samples using Illumina HiSeq2500 by a blinded analyst. RESULTS All human tumors (n=7) displayed region-specific long projecting tracts extending into brain parenchyma (Mean=23.2mm, SD=3.1mm). Maximum tract length varied from 80–130mm (Mean=102mm, SD=20.4mm). Xenograft models displayed similar tumor-intersecting tractography (n=3), with 3D reconstruction of stained GSCs replicating that pattern. RNAseq data revealed significant overrepresentation (≥2-fold) of 528 transcripts in projecting tumor tract-associated samples compared to samples obtained from the tumor mass itself and brain parenchyma unassociated with projecting tumor tracts. Functional classification revealed that 62% of these transcripts regulate cell motility as part of inter-related networks. CONCLUSION These imaging data, backed by region-specific transcriptomic evidence, suggest that Q-Space MRI may discriminate localizable patterns of inter-voxel coherence representing tumor-associated infiltration pathways in glioblastoma.

Assessment of transverse ligament (TL) competence in patients with suspected atlantoaxial instability is performed via indirect radiograph measurements or direct TL visualization on magnetic resonance imaging (MRI). Interpretation of these images can be limited by unique patient anatomy or imaging technique variability. We report a novel technique for evaluating TL competence using flexion-extension computed tomography (feCT) scan with 3-dimensional (3D) segmentation and quantitative analysis.feCT scans of 11 patients were segmented to create 3D surface models. Six patients with atlantoaxial pathology were evaluated for possible instability based on clinical examination and imaging findings. The other 5 patients had no clinical or imaging evidence of atlantoaxial injury. Dynamic atlantodental interval (ADI) was calculated using point-to-point voxel changes between flexion and extension 3D models. Magnitude and direction of ADI changes were quantified and compared with available cervical spine flexion-extension radiograph and/or MRI findings.In the 5 patients without evidence of atlantoaxial injury, 94.3% of ADI vector changes were3.0 mm. In the 3 patients with atlantoaxial pathology but TL competence, 92.4% of ADI vector changes were3.0 mm. In the 3 patients with atlantoaxial pathology and TL incompetence, only 49.1% of ADI vector changes were3.0 mm. In addition to the significant atlantoaxial subluxation in these 3 patients, there was significant rotational motion compared with the patients with an intact TL.3D segmentation and quantitative analysis of feCT scan allow objective indirect assessment of TL integrity. Results are consistent with MRI findings and offer additional biomechanical information regarding the direction and distribution of atlantoaxial motion.

Objective: Investigators conducted pilot sessions in live Emergency Department (ED) settings to prepare for formal comparative study of contact and experimental non-contact vital signs (VS) measurement devices. Methods: Contact-based cardiorespiratory monitors (CM) and video photoplethysmography (vPPG) measured heart rates (HR); contact thermometers and passive infrared thermography (pIR) obtained core and surface temperatures. Subject VS data comprised two 25-min CM and vPPG recordings (initial unprimed “Triage Check” window; four 1-min “Full Check” windows; 20min of “Resting” intervals) and temperature measurements. Results: Forty-eight of 102 approached patients participated: at least 5 subjects each in 0-12mo, 1-5y, 6-12y, 13-17y age-groups, and at least 2 subjects each in 18-29y, 30-39y, 40-49y, 50-59y, 60-69y, 70-79y, 80+y age-groups. Subjects were 40.4% female with median Fitzpatrick skin type of 3.0 (interquartile range 1-3: 2.0-5.0), ESI score of 3.0 (2.0-3.0), and core temperature of 98.6°F (97.9-99.1°F). Twelve subjects were excluded from vPPG analysis due to inadequate lighting, excessive motion, and/or datastream loss. From the remaining 36 subjects with median HR CM of 78.2bpm (69.8-94.8bpm), vPPG measured median HR vPPG of 75.0bpm (68.3-93.0bpm) from 170 (70.5%) of 241 possible Full Check windows; the median difference between HR CM and HR vPPG was 0.6bpm (-0.2-2.3bpm). All Emergency Severity Index (ESI) HR components for 34 subjects with Triage Check HR vPPG measurements matched CM-based ESI HR components. Median differences between contact and pIR temperatures ranged from −5.4°F to −4.4°F. Conclusion: vPPG and pIR devices experimentally measured select live VS with promising results. Significance: Pilot data and protocol testing set the groundwork for full-scale ED vPPG-pIR investigation.

Aspartate-glutamate carrier 1 (AGC1) is one of two exchangers within the malate-aspartate shuttle. AGC1 is encoded by the SLC25A12 gene. Three patients with pathogenic variants in SLC25A12 have been reported in the literature. These patients were clinically characterized by neurodevelopmental delay, epilepsy, hypotonia, cerebral atrophy, and hypomyelination; however, there has been discussion in the literature as to whether this hypomyelination is primary or secondary to a neuronal defect. Here we report a 12-year-old patient with variants in SLC25A12 and magnetic resonance imaging (MRI) at multiple ages. Novel compound heterozygous, recessive variants in SLC25A12 were identified: c.1295C>T (p.A432V) and c.1447-2_1447-1delAG. Clinical presentation is characterized by severe intellectual disability, nonambulatory, nonverbal status, hypotonia, epilepsy, spastic quadriplegia, and a happy disposition. The serial neuroimaging findings are notable for cerebral atrophy with white matter involvement, namely, early hypomyelination yet subsequent progression of myelination. The longitudinal MRI findings are most consistent with a leukodystrophy of the leuko-axonopathy category, that is, white matter abnormalities that are most suggestive of mechanisms that result from primary neuronal defects. We present here the first case of a patient with compound heterozygous variants in SLC25A12, including brain MRI findings, in the oldest individual reported to date with this neurogenetic condition.

Our objective is to evaluate the effectiveness of efficient convolutional neural networks (CNNs) for abnormality detection in chest radiographs and investigate the generalizability of our models on data from independent sources. We used the National Institutes of Health ChestX-ray14 (NIH-CXR) and the Rhode Island Hospital chest radiograph (RIH-CXR) datasets in this study. Both datasets were split into training, validation, and test sets. The DenseNet and MobileNetV2 CNN architectures were used to train models on each dataset to classify chest radiographs into normal or abnormal categories; models trained on NIH-CXR were designed to also predict the presence of 14 different pathological findings. Models were evaluated on both NIH-CXR and RIH-CXR test sets based on the area under the receiver operating characteristic curve (AUROC). DenseNet and MobileNetV2 models achieved AUROCs of 0.900 and 0.893 for normal versus abnormal classification on NIH-CXR and AUROCs of 0.960 and 0.951 on RIH-CXR. For the 14 pathological findings in NIH-CXR, MobileNetV2 achieved an AUROC within 0.03 of DenseNet for each finding, with an average difference of 0.01. When externally validated on independently collected data (e.g., RIH-CXR-trained models on NIH-CXR), model AUROCs decreased by 3.6–5.2% relative to their locally trained counterparts. MobileNetV2 achieved comparable performance to DenseNet in our analysis, demonstrating the efficacy of efficient CNNs for chest radiograph abnormality detection. In addition, models were able to generalize to external data albeit with performance decreases that should be taken into consideration when applying models on data from different institutions.

Background: White matter disease (WMD) and microhemorrhages (MH) present at the time of stroke have been linked to outcome, yet few have investigated changes in the time leading up to stroke. Magnetic Resonance Imaging (MRI) characteristics before a stroke may shed light on the severity of outcomes following stroke. Methods: We retrospectively identified patients admitted to our institution for ischemic stroke between 5/16-12/17 who had an MRI in the 1-24 months prior to their stroke. After collecting clinical and demographic data, automatic segmentations of gray and white matter volumes and manual segmentation of WMD and MH (size Results: Among the 48 patients in our study (mean age 68, 50% female), 29 patients (60%) had poor outcome. Poor outcome was associated with pre-existing WMD volume (21.9+/-23.6 vs 6.1+/-5.7mL, p= 0.002), and WMD volume at time of stroke (26.4+/-24.9 vs 8.8+/-6.0mL, p=0.001). There was a trend towards poor outcome in older patients (p=0.073), women (p=0.075), higher NIHSS score (p=0.066), and need for thrombectomy (p=0.065). History of prior stroke was not associated with poor outcome (p=0.74), larger pre-existing WMD volumes (14.3+/-19.3 vs 19.5+/-22.6mL, p=0.47) or larger WMD volumes at time of stroke (18.7±22.5 vs 21.5±18.8mL, p=0.67). After adjusting for confounders, pre-existing WMD volume showed a trend to predict poor outcome (adjusted OR 1.086 per one point increase, 95% CI 0.987-1.195, p=0.09), as did WMD volume at the time of stroke (adjusted OR 1.103 per one point increase, 95% CI 0.992-1.226, p=0.07). Conclusions: Greater volume of WMD at time of and prior to admission for stroke may be independent predictors of poor outcome. These results should be validated in subsequent studies.

Background: Patients are more likely to develop white matter disease (WMD) and microhemorrhage (MH) after ischemic stroke. Little is known about whether stroke patients had an increased prevalence of these findings before a stroke compared to patients who did not go on to develop a stroke. Methods: We retrospectively identified patients admitted to our institution with ischemic stroke between 5/16 - 12/17 who had magnetic resonance imaging (MRI) of the brain between 30 days-2 years prior to their stroke. Age and gender-matched controls for the initial MRI were identified for pair-wise comparison. Automatic segmentations of gray and white matter volumes and manual segmentation of WMD and MH (size Results: Among the 96 patients in our study, stroke patients (n=48, mean age 68, 50% female) had lower rates of cancer (p=0.03) but higher rates of hyperlipidemia (p=0.024) compared to controls. Stroke patients exhibited greater WMD volume (15.7±20.2 vs 7.3±10.4 mL, p= 0.012) compared to controls, but not MH number (p=0.29) or total MH volume (p=0.26). Although history of stroke was associated with new presentation of stroke (72.9% vs 22.9%, p Conclusion: Increased volume of white matter disease correlates with a greater risk of future ischemic stroke. These results should be validated in subsequent studies.

INTRODUCTION Effective resuscitation of out-of-hospital cardiac arrest (OHCA) patients is challenging. Alternative resuscitative approaches using electromechanical adjuncts may improve provider performance. Investigators applied simulation to study the effect of an experimental automation-assisted, goal-directed OHCA management protocol on EMS providers' resuscitation performance relative to standard protocols and equipment. METHODS Two-provider (emergency medical technicians (EMT)-B and EMT-I/C/P) teams were randomized to control or experimental group. Each team engaged in 3 simulations: baseline simulation (standard roles); repeat simulation (standard roles); and abbreviated repeat simulation (reversed roles, i.e., basic life support provider performing ALS tasks). Control teams used standard OHCA protocols and equipment (with high-performance cardiopulmonary resuscitation training intervention); for second and third simulations, experimental teams performed chest compression, defibrillation, airway, pulmonary ventilation, vascular access, medication, and transport tasks with goal-directed protocol and resuscitation-automating devices. Videorecorders and simulator logs collected resuscitation data. RESULTS Ten control and 10 experimental teams comprised 20 EMT-B's; 1 EMT-I, 8 EMT-C's, and 11 EMT-P's; study groups were not fully matched. Both groups suboptimally performed chest compressions and ventilations at baseline. For their second simulations, control teams performed similarly except for reduced on-scene time, and experimental teams improved their chest compressions (P=0.03), pulmonary ventilations (P

For computational models of microwave ablation (MWA), knowledge of the antenna design is necessary, but the proprietary design of clinical applicators is often unknown. We characterised the specific absorption rate (SAR) during MWA experimentally and compared to a multi-physics simulation.An infrared (IR) camera was used to measure SAR during MWA within a split ex vivo liver model. Perseon Medical's short-tip (ST) or long-tip (LT) MWA antenna were placed on top of a tissue sample (n = 6), and microwave power (15 W) was applied for 6 min, while intermittently interrupting power. Tissue surface temperature was recorded via IR camera (3.3 fps, 320 × 240 resolution). SAR was calculated intermittently based on temperature slope before and after power interruption. Temperature and SAR data were compared to simulation results.Experimentally measured SAR changed considerably once tissue temperatures exceeded 100 °C, contrary to simulation results. The ablation zone diameters were 1.28 cm and 1.30 ± 0.03 cm (transverse), and 2.10 cm and 2.66 ± -0.22 cm (axial), for simulation and experiment, respectively. The average difference in temperature between the simulation and experiment were 5.6 °C (ST) and 6.2 °C (LT). Dice coefficients for 1000 W/kg SAR iso-contour were 0.74 ± 0.01 (ST) and 0.77 (± 0.03) (LT), suggesting good agreement of SAR contours.We experimentally demonstrated changes in SAR during MWA ablation, which were not present in simulation, suggesting inaccuracies in dielectric properties. The measured SAR may be used in simplified computer simulations to predict tissue temperature when the antenna geometry is unknown.

Purpose: The objective of this study is to measure through simulation the impact of (1) heterogeneity of biophysical parameters in tumor vs healthy tissue, (2) applicator placement relative to the tumor, and (3) proximity to large blood vessels on microwave ablation (MWA) treatment effect area. This will help identify the biophysical properties that have the greatest impact on improving clinical modeling of MWA procedures. Methods: The authors’ approach was to develop two-compartment models with variable tissue properties and simulate MWA procedures performed in liver with Perseon Medical’s 915 MHz short-tip applicator. Input parameters for the dielectric and thermal properties considered in this study were based on measurements for healthy and malignant (primary or metastatic) liver tissue previously reported in the literature. Compartment 1 (C1) represented normal, fatty, or cirrhotic liver, and compartment 2 (C2) represented a primary hepatocellular carcinomatumor sample embedded within C1. To evaluate the sensitivity to tissue parameters, a range of clinically relevant tissue properties were simulated. To evaluate the impact of MWA antenna position, the authors simulated various tumor perfusion models with the antenna shifted 5 mm anteriorly and posteriorly. To evaluate the effect of local vasculature, the authors simulated an additional heat sink of various diameters and distances from the tumor. Dice coefficient statistics were used to evaluate ablation zone effects from these local heat sinks. Results: Models showed less than 11% of volume variability (1 cm3 increase) in ablation treatment effect region when accounting for the difference in relative permittivity and electrical conductivity between malignant and healthy liver tissue. There was a 27% increase in volume when simulating thermal conductivity of fatty liver disease versus the baseline simulation. The ablation zone volume increased more than 36% when simulating cirrhotic surrounding liver tissue.Antenna placement relative to the tumor had minimal sensitivity to the absolute size of the treatment effect area, with less than 1.5 mm variation. However, when considering the overlap between the ablation zone and the ideal clinical margin when the antenna was displaced 5 mm anteriorly and posteriorly, there was approximately a 6 mm difference in the margins. Dice coefficient statistics showed as much as an 11% decrease in the ablation margin due to the presence of vessel heat sinks within the model. Conclusions: The results from simulating the variance in malignant tissue thermal and electrical properties will help guide better approximations for MWA treatments. The results suggest that assuming malignant and healthy liver tissues have similar dielectric properties is a reasonable first approximation. Antenna placement relative to the tumor has minimal impact on the absolute size of the ablation zone, yet it does cause relevant variation between desired treatment margin and ablation zone. Blood vessel cooling, especially hepatic vessels close to the region of interest, may be a significant factor to consider in treatment planning. Further data need to be collected for assessing treatment planning utility of modeling MWA in this context.

PURPOSE: To develop a deep learning approach to bone age assessment based on a training set of developmentally normal pediatric hand radiographs and to compare this approach with automated and manual bone age assessment methods based on Greulich and Pyle (GP). METHODS: In this retrospective study, a convolutional neural network (trauma hand radiograph–trained deep learning bone age assessment method [TDL-BAAM]) was trained on 15 129 frontal view pediatric trauma hand radiographs obtained between December 14, 2009, and May 31, 2017, from Children’s Hospital of New York, to predict chronological age. A total of 214 trauma hand radiographs from Hasbro Children’s Hospital were used as an independent test set. The test set was rated by the TDL-BAAM model as well as a GP-based deep learning model (GPDL-BAAM) and two pediatric radiologists (radiologists 1 and 2) using the GP method. All ratings were compared with chronological age using mean absolute error (MAE), and standard concordance analyses were performed. RESULTS: The MAE of the TDL-BAAM model was 11.1 months, compared with 12.9 months for GPDL-BAAM (P = .0005), 14.6 months for radiologist 1 (P < .0001), and 16.0 for radiologist 2 (P < .0001). For TDL-BAAM, 95.3% of predictions were within 24 months of chronological age compared with 91.6% for GPDL-BAAM (P = .096), 86.0% for radiologist 1 (P < .0001), and 84.6% for radiologist 2 (P < .0001). Concordance was high between all methods and chronological age (intraclass coefficient > 0.93). Deep learning models demonstrated a systematic bias with a tendency to overpredict age for younger children versus radiologists who showed a consistent mean bias. CONCLUSION: A deep learning model trained on pediatric trauma hand radiographs is on par with automated and manual GP-based methods for bone age assessment and provides a foundation for developing population-specific deep learning algorithms for bone age assessment in modern pediatric populations. Supplemental material is available for this article. © RSNA, 2020 See also the commentary by Halabi in this issue.

Introduction Protocolized automation of critical, labor-intensive tasks for out-of-hospital cardiac arrest (OHCA) resuscitation may decrease Emergency Medical Services (EMS) provider workload. A simulation-based assessment method incorporating objective and self-reported metrics was developed and used to quantify workloads associated with standard and experimental approaches to OHCA resuscitation. Methods Emergency Medical Services-Basic (EMT-B) and advanced life support (ALS) providers were randomized into two-provider mixed-level teams and fitted with heart rate (HR) monitors for continuous HR and energy expenditure (EE) monitoring. Subjects' resting salivary α-amylase (sAA) levels were measured along with Borg perceived exertion scores and multidimensional workload assessments (NASA-TLX). Each team engaged in the following three OHCA simulations: (1) baseline simulation in standard BLS/ALS roles; (2) repeat simulation in standard roles; and then (3) repeat simulation in reversed roles, ie, EMT-B provider performing ALS tasks. Control teams operated with standard state protocols and equipment; experimental teams used resuscitation-automating devices and accompanying goal-directed algorithmic protocol for simulations 2 and 3. Investigators video-recorded resuscitations and analyzed subjects' percent attained of maximal age-predicted HR (%mHR), EE, sAA, Borg, and NASA-TLX measurements. Results Ten control and ten experimental teams completed the study (20 EMT-Basic; 1 EMT-Intermediate, 8 EMT-Cardiac, 11 EMT-Paramedic). Median %mHR, EE, sAA, Borg, and NASA-TLX scores did not differ between groups at rest. Overall multivariate analyses of variance did not detect significant differences; univariate analyses of variance for changes in %mHR, Borg, and NASA-TLX from resting state detected significant differences across simulations (workload reductions in experimental groups for simulations 2 and 3). Conclusions A simulation-based OHCA resuscitation performance and workload assessment method compared protocolized automation-assisted resuscitation with standard response. During exploratory application of the assessment method, subjects using the experimental approach appeared to experience reduced levels of physical exertion and perceived workload than control subjects.

Accurate non-contact acquisition of patient vital signs will advance emergency care. In order to assess promising candidate technologies., an observational study was conducted with healthy volunteers to test two hypotheses: 1. Video photoplethysmography and motion analysis (vPPG-MA) and infrared thermography (IR) will accurately and concurrently measure heart rate (HR) and respiratory rate (RR)., and body temperature., respectively. 2. Non-contact approaches will exhibit comparable and reliable performance against standard contact cardiorespiratory monitors (CM). HR and RR were measured with CM and vPPG-MA; core and surface temperatures were obtained using oral thermometry and two IR cameras., respectively. Subjects were videorecorded at rest; during sustained exercise at 50%., 60%., and 70% of age-predicted maximum HR; and 1., 3., and 5 min post-exertion. vPPG-MA HR and RR measurements were calculated for video segments corresponding to ED use-cases: Triage (unprimed) 30s check., Routine 30s check, Abbreviated “Spot” 10s check., and Full 60s check. Descriptive statistics and Bland-Altman analyses were performed on vPPG-MA and IR measurements against synchronous CM measurements. Thirty volunteers exhibited a HR range of 43-146bpm., a RR range of 8-29bpm., and an oral temperature range of 96.2-99.5°F on CM. vPPG-MA obtained 972 (98.2% of scheduled) HR and 591 (98.5%) RR measurements; mean differences between Full 60s vPPG-MA and CM were −0.9±5.5bpm (-0.9±5.3%; 95% CI: −11.6–9.8bpm) for HR, and 0.9±3.1bpm (4.8±17.6%; −5.1–6.9bpm) for RR; other video segments performed similarly. IR acquired temperatures ~4°F lower than oral thermometers. vPPG-MA and IR thermography successfully measured select vital signs concurrently. vPPG-MA‘s observed level of agreement with CM, along with temperature offsets identified for IR-based thermometry., have set the foundation for live ED clinical studies.

A significant limitation of minimally invasive surgery is dependence of the entire surgical team on a single endoscopic viewpoint. An individualized, instrument-driven image display system that allows all operators to simultaneously define their viewing frame of the surgical field may be the solution. We tested the efficacy of such a system using a modified Fundamentals of Laparoscopic Surgery™ (Society of American Gastrointestinal and Endoscopic Surgeons, Los Angeles, CA) bead transfer task.A program was custom-written to allow zooming and centering of the image window on specific color signals, each attached near the tip of a different laparoscopic instrument. Two controls were used for the bead transfer task: (1) a static, wide-angle view and (2) a single moving camera allowing close-up and tracking of the bead as it was transferred. Time to task completion and number of bead drops were recorded.Thirty-six sessions were performed by surgical residents. Average time for bead transfer was 127.3±21.3 seconds in the Experimental group, 139.1±27.8 seconds in the Control 1 group, and 186.2±18.5 seconds in the Control 2 group (P=.034, by analysis of variance). Paired analysis (the Wilcoxon Signed-Rank Test) showed that the Experimental group was significantly faster than the Control 1 group (P=.035) and the Control 2 group (P=.028).We have developed an image navigation system that allows intuitive and efficient laparoscopic performance compared with two controls. It offers high-resolution images and ability for multitasking. The tracking system centers close-up images on the laparoscopic target. Further development of robust prototypes will help transition this in vitro system into clinical application.

The combination of healing anterior cruciate ligament (ACL) volume and the distributions of T 2 ⁎ relaxation times within it have been shown to predict the biomechanical failure properties in a porcine model. This MR-based prediction model has not yet been used to assess ligament degeneration in the aging human knee. Using a set of 15 human cadaveric knees of varying ages, we obtained in situ MR measures of volume and T 2 ⁎ of the intact ACL and then related these MR variables to biomechanical outcomes (maximum and yield loads, linear stiffness) obtained via ex vivo failure testing. Using volume in conjunction with the median T 2 ⁎ value, the multiple linear regression model did not predict maximum failure load for the intact human ACL; R2=0.23, p=0.200. Similar insignificant results were found for yield load and linear stiffness. Naturally restricted distributions of the intact ligament volume and T 2 ⁎ (demonstrated by the respective Z-scores) in an older cadaveric population were the likely reason for the insignificant results. These restricted distributions may negatively affect the ability to detect a correlation when one exists. Further research is necessary to understand the relationship of MRI variables and ligament degeneration. While this study failed to find a significant prediction of human biomechanical outcome using these MR variables, with further research, an MR-based approach may offer a tool to longitudinally assess changes in cruciate ligament degradation.

Author(s): Kobayashi, Leo; Zhang, Xiao Chi; Collins, Scott A.; Karim, Naz; Merck, Derek L. | Abstract: Introduction: Augmented reality (AR), mixed reality (MR), and virtual reality devices are enabling technologies that may facilitate effective communication in healthcare between those with information and knowledge (clinician/specialist; expert; educator) and those seeking understanding and insight (patient/family; non-expert; learner). Investigators initiated an exploratory program to enable the study of AR/MR use-cases in acute care clinical and instructional settings.Methods: Academic clinician educators, computer scientists, and diagnostic imaging specialists conducted a proof-of-concept project to 1) implement a core holoimaging pipeline infrastructure and open-access repository at the study institution, and 2) use novel AR/MR techniques on off-the-shelf devices with holoimages generated by the infrastructure to demonstrate their potential role in the instructive communication of complex medical information.Results: The study team successfully developed a medical holoimaging infrastructure methodology to identify, retrieve, and manipulate real patients’ de-identified computed tomography and magnetic resonance imagesets for rendering, packaging, transfer, and display of modular holoimages onto AR/MR headset devices and connected displays. Holoimages containing key segmentations of cervical and thoracic anatomic structures and pathology were overlaid and registered onto physical task trainers for simulation-based “blind insertion” invasive procedural training. During the session, learners experienced and used task-relevant anatomic holoimages for central venous catheter and tube thoracostomy insertion training with enhanced visual cues and haptic feedback. Direct instructor access into the learner’s AR/MR headset view of the task trainer was achieved for visual-axis interactive instructional guidance.Conclusion: Investigators implemented a core holoimaging pipeline infrastructure and modular open-access repository to generate and enable access to modular holoimages during exploratory pilot stage applications for invasive procedure training that featured innovative AR/MR techniques on off-the-shelf headset devices.

The purpose of our study was to determine if a textural analysis metric can be implemented to improve diagnosis of adenomyosis by ultrasound.We retrospectively identified 38 patients with a magnetic resonance imaging (MRI) diagnosis of uterine adenomyosis that also had a pelvic ultrasound within 6 months. We also identified 50 normal pelvic ultrasound examinations confirmed by a normal pelvic MRI within 6 months as a control group. A region of interest (ROI) was subsequently placed on the study population ultrasound image corresponding to the area of adenomyosis on MRI. An ROI was placed in the area of the junctional zone in the normal controls. The abnormal and normal ROIs were then compared against trained normal and abnormal distributions to determine the success rate, sensitivity, specificity, and negative and positive predictive values of our computer metric. The ultrasound reports performed before MRI were also reviewed to determine the radiologist correct/incorrect interpretation rate for comparison with our textural analysis metric.Using a training population of 50 normal ultrasound examinations (confirmed with a normal MRI) and 38 abnormal ultrasound examinations (MRI confirmed adenomyosis), we had an overall 75% (66/88 accurately diagnosed) success rate with a sensitivity, specificity, and negative and positive predictive values of 70%, 79%, 73%, and 76%, respectively (P < .0001). The sensitivity and false-negative rate of the initial ultrasound interpretation were 26% (10/38) and 74% (28/38), respectively.

Objectives: (1) To develop a clinical microsystem simulation methodology for alarm fatigue research with a human factors engineering (HFE) assessment framework and (2) to explore its application to the comparative examination of different approaches to patient monitoring and provider notification. Background: Problems with the design, implementation, and real-world use of patient monitoring systems result in alarm fatigue. A multidisciplinary team is developing an open-source tool kit to promote bedside informatics research and mitigate alarm fatigue. Method: Simulation, HFE, and computer science experts created a novel simulation methodology to study alarm fatigue. Featuring multiple interconnected simulated patient scenarios with scripted timeline, “distractor” patient care tasks, and triggered true and false alarms, the methodology incorporated objective metrics to assess provider and system performance. Developed materials were implemented during institutional review board–approved study sessions that assessed and compared an experimental multiparametric alerting system with a standard monitor telemetry system for subject response, use characteristics, and end-user feedback. Results: A four-patient simulation setup featuring objective metrics for participant task-related performance and response to alarms was developed along with accompanying structured HFE assessment (questionnaire and interview) for monitor systems use testing. Two pilot and four study sessions with individual nurse subjects elicited true alarm and false alarm responses (including diversion from assigned tasks) as well as nonresponses to true alarms. In-simulation observation and subject questionnaires were used to test the experimental system’s approach to suppressing false alarms and alerting providers. Conclusions: A novel investigative methodology applied simulation and HFE techniques to replicate and study alarm fatigue in controlled settings for systems assessment and experimental research purposes.

The objective of this study is to measure through simulation the impact of (1) heterogeneity of biophysical parameters in tumor vs healthy tissue, (2) applicator placement relative to the tumor, and (3) proximity to large blood vessels on microwave ablation (MWA) treatment effect area. This will help identify the biophysical properties that have the greatest impact on improving clinical modeling of MWA procedures.The authors' approach was to develop two-compartment models with variable tissue properties and simulate MWA procedures performed in liver with Perseon Medical's 915 MHz short-tip applicator. Input parameters for the dielectric and thermal properties considered in this study were based on measurements for healthy and malignant (primary or metastatic) liver tissue previously reported in the literature. Compartment 1 (C1) represented normal, fatty, or cirrhotic liver, and compartment 2 (C2) represented a primary hepatocellular carcinoma tumor sample embedded within C1. To evaluate the sensitivity to tissue parameters, a range of clinically relevant tissue properties were simulated. To evaluate the impact of MWA antenna position, the authors simulated various tumor perfusion models with the antenna shifted 5 mm anteriorly and posteriorly. To evaluate the effect of local vasculature, the authors simulated an additional heat sink of various diameters and distances from the tumor. Dice coefficient statistics were used to evaluate ablation zone effects from these local heat sinks.Models showed less than 11% of volume variability (1 cm(3) increase) in ablation treatment effect region when accounting for the difference in relative permittivity and electrical conductivity between malignant and healthy liver tissue. There was a 27% increase in volume when simulating thermal conductivity of fatty liver disease versus the baseline simulation. The ablation zone volume increased more than 36% when simulating cirrhotic surrounding liver tissue. Antenna placement relative to the tumor had minimal sensitivity to the absolute size of the treatment effect area, with less than 1.5 mm variation. However, when considering the overlap between the ablation zone and the ideal clinical margin when the antenna was displaced 5 mm anteriorly and posteriorly, there was approximately a 6 mm difference in the margins. Dice coefficient statistics showed as much as an 11% decrease in the ablation margin due to the presence of vessel heat sinks within the model.The results from simulating the variance in malignant tissue thermal and electrical properties will help guide better approximations for MWA treatments. The results suggest that assuming malignant and healthy liver tissues have similar dielectric properties is a reasonable first approximation. Antenna placement relative to the tumor has minimal impact on the absolute size of the ablation zone, yet it does cause relevant variation between desired treatment margin and ablation zone. Blood vessel cooling, especially hepatic vessels close to the region of interest, may be a significant factor to consider in treatment planning. Further data need to be collected for assessing treatment planning utility of modeling MWA in this context.

Learning probability distributions of the shape of anatomic structures requires fitting shape representations to human expert segmentations from training sets of medical images. The quality of statistical segmentation and registration methods is directly related to the quality of this initial shape fitting, yet the subject is largely overlooked or described in an ad hoc way. This article presents a set of general principles to guide such training. Our novel method is to jointly estimate both the best geometric model for any given image and the shape distribution for the entire population of training images by iteratively relaxing purely geometric constraints in favor of the converging shape probabilities as the fitted objects converge to their target segmentations. The geometric constraints are carefully crafted both to obtain legal, nonself-interpenetrating shapes and to impose the model-to-model correspondences required for useful statistical analysis. The paper closes with example applications of the method to synthetic and real patient CT image sets, including same patient male pelvis and head and neck images, and cross patient kidney and brain images. Finally, we outline how this shape training serves as the basis for our approach to IGRT∕ART.

We propose an adaptable framework for analyzing ultrasound (US) images quantitatively to provide computer-aided diagnosis using machine learning. Our preliminary clinical targets are hepatic steatosis, adenomyosis, and craniosynostosis. For steatosis and adenomyosis, we collected US studies from 288 and 88 patients, respectively, as well as their biopsy or magnetic resonanceconfirmed diagnosis. Radiologists identified a region of interest (ROI) on each image. We filtered the US images for various texture responses and use the pixel intensity distribution within each ROI as feature parameterizations. Our craniosynostosis dataset consisted of 22 CT-confirmed cases and 22 age-matched controls. One physician manually measured the vectors from the center of the skull to the outer cortex at every 10deg for each image and we used the principal directions as shape features for parameterization. These parameters and the known diagnosis were used to train classifiers. Testing with cross-validation, we obtained 72.74% accuracy and 0.71 area under receiver operating characteristics curve for steatosis ([Formula: see text]), 77.27% and 0.77 for adenomyosis ([Formula: see text]), and 88.63% and 0.89 for craniosynostosis ([Formula: see text]). Our framework is able to detect a variety of diseases with high accuracy. We hope to include it as a routinely available support system in the clinic.