Butterfly Network Research Articles

Point-of-Care Ultrasound on Clinical Rotations: Reported Opportunity and Objective Number of Examinations

Introduction Point-of-care ultrasound (POCUS) inclusion in physician assistant (PA) program curricula is increasing. Research has shown that preceptors' readiness to teach POCUS on rotations is limited. The purpose of this research is to describe PA students' self-reported opportunity to practice POCUS on rotations, to quantify the number of POCUS examinations completed on rotations, and to analyze data to provide insights into which rotations might be most amendable to POCUS training for PA students. Methods Students were surveyed about their preceptors' willingness to allow the practice of POCUS on rotations. Students were required to complete POCUS examinations during rotations. The number of examinations performed was extracted from the Butterfly Network software and paired with the students' rotation information. Descriptive statistics were completed. Results A survey response rate of 71.4% was noted. The student reported opportunities to practice POCUS were the highest on Family Medicine, Cardiology, Hospital Internal Medicine (HIM), and Emergency Medicine (EM). Students completed 778 POCUS examinations on rotations. A higher objective number of examinations were completed on EM and HIM rotations. Discussion EM and HIM rotations afforded the most opportunity to convert the reported opportunity to practice POCUS to the objective act of obtaining POCUS images. Most rotations had proportionally similar number of POCUS examinations completed per organ system, however, Cardiology had a disproportionately higher number of cardiac examinations. This rotation and application-specific understanding of PA students' ability to obtain POCUS examinations on rotations should be used by PA programs while developing POCUS curriculum.

Comparison of 6 handheld ultrasound devices by point-of-care ultrasound experts: a cross-sectional study

BackgroundPoint-of-care ultrasound (POCUS) has emerged as an essential bedside tool for clinicians, but lack of access to ultrasound equipment has been a top barrier to POCUS use. Recently, several handheld ultrasound devices (“handhelds”) have become available, and clinicians are seeking data to guide purchasing decisions. Few comparative studies of different handhelds have been done. We conducted a cross-sectional study comparing 6 handhelds readily available in the United States (Butterfly iQ + ™ by Butterfly Network Inc.; Clarius™ by Clarius Mobile Health; Kosmos™ by EchoNous; TE Air™ by Mindray; Vscan Air™ SL and CL by General Electric; and Lumify™ by Philips Healthcare). A multi-specialty group of physician POCUS experts (n = 35) acquired three standard ultrasound views (abdominal right upper quadrant, cardiac apical 4-chamber, and superficial neck and lung views) in random order on the same standardized patients and rated the image quality. Afterward, a final survey of the overall ease of use, image quality, and satisfaction of each handheld was completed.ResultsThirty-five POCUS experts specializing in internal medicine/hospital medicine, critical care, emergency medicine, and nephrology acquired and rated right upper quadrant, apical 4-chamber, and superficial neck and lung views with 6 different handhelds. For image quality, the highest-rated handhelds were Vscan Air™ for the right upper quadrant view, Mindray TE Air™ for the cardiac apical 4-chamber view, and Lumify™ for superficial views of the neck and lung. Overall satisfaction with image quality was highest with Vscan Air™, Lumify™, and Mindray, while overall satisfaction with ease of use was highest with Vscan Air™. The 5 most desirable characteristics of handhelds were image quality, ease of use, portability, probe size, and battery life. Ultimately, all 6 handhelds had notable advantages and disadvantages, with no single device having all desired qualities or features.ConclusionsThe overall satisfaction with image quality was rated highest with Vscan Air™, Lumify™, and Mindray TE Air™when acquiring right upper quadrant, apical 4-chamber, and superficial neck and lung views. No single handheld was perceived to be superior in image quality for all views. Vscan Air™ was rated highest for overall ease of use and was the most preferred handheld for purchase by POCUS experts.

On the Geodetic and Strong Geodetic Parameters of Certain Hypercubic Networks

Networks derived from the hypercube are noted for their efficiency in simulating any bounded-degree communication network and hence they play a significant role in parallel machines. For a graph [Formula: see text], a geodetic cover is a subset [Formula: see text] of [Formula: see text], satisfying the condition that every vertex of [Formula: see text] other than the vertices in [Formula: see text] are covered using geodesics connecting any two vertices in [Formula: see text]. The geodetic number of [Formula: see text] is the least order of such covers in [Formula: see text]. This seminal distance-based parameter is well employed in the areas of location theory, convexity theory and game theory. The strong variant and its edge version are recent variations of the problem that were introduced in due of their applications in social networks. The most established and strong hypercube derivative networks are studied in this paper with respect to certain geodetic parameters. Notably, among the networks that we have considered, the butterfly and the Benes networks belong to the class of bipartite graphs for which the geodetic number problem and its strong version are NP-complete. We have determined the geodetic parameters that include the geodetic number, the strong geodetic number and their corresponding edge variants of the networks taken into consideration. In the process, we have identified a new class of graphs, as semi-extreme edge geodesic graphs and that the enhanced butterfly networks belong to this class. The study of such geodetic parameters in the structural models of parallel computation could be useful for the design of efficient algorithms for the structures considered.

Prior image-based generative adversarial learning for multi-material decomposition in photon counting computed tomography

BackgroundPhoton counting detector computed tomography (PCD-CT) is a novel promising technique providing higher spatial resolution, lower radiation dose and greater energy spectrum differentiation, which create more possibilities to improve image quality. Multi-material decomposition is an attractive application for PCD-CT to identify complicated materials and provide accurate quantitative analysis. However, limited by the finite photon counting rate in each energy window of photon counting detector, the noise problem hinders the decomposition of high-quality basis material images. MethodsTo address this issue, an end-to-end multi-material decomposition network based on prior images is proposed in this paper. First, the reconstructed images corresponding to the full spectrum with less noise are introduced as prior information to improve the overall signal-to-noise ratio of the data. Then, a generative adversarial network is designed to mine the relationship between reconstructed images and basis material images based on the information interaction of material decomposition. Furthermore, a weighted edge loss is introduced to adapt to the structural differences of different basis material images. ResultsTo verify the performance of the proposed method, simulation and real studies are carried out. In simulation study of structured fibro-glandular tissue model, the results show that the proposed method decreased the root mean square error by 67 % and 26 % on adipose, 66 % and 28 % on fibroglandular, 52 % and 8 % on calcification, compared to butterfly network and dual interactive Wasserstein generative adversarial network. ConclusionExperimentally, the proposed method shows certain advantages over other methods on noise suppression effect, detail retention ability and decomposition accuracy.

Cross quantum homomorphic encryption based on the butterfly network

Cross quantum homomorphic encryption based on the butterfly network

Fuzzy Tree t -Spanners of Ratio Labelled Butterfly Network

Background/Objectives: The reliability and optimality of a communication system depend on the network that has a minimum distance. The minimum distance is the shortest path between the vertices. Even in error correction codes, the minimum Hamming distance ensures data integrity. Hence to have efficient routing protocols, the network is preferred to be a spanning tree that connects all the nodes of the communication system. As tree t- spanner is the optimization technique in determining minimum spanning tree, finding tree t-spanners of a ratio labelled fuzzy graph for the Butterfly network BF(n) is notion of this study. Methods: This study has introduced a new labelling called Ratio Labelling (RL) for examining the fuzziness in a crisp graph. Using the RL method, the crisp graphs were examined for the admittance of fuzziness and were thereby found, to have a tree t-spanners of ratio labelled fuzzy graphs. Findings: The quality of a network can be measured using the stretch factor of the network. The issues of network survival or link failures in a communication networks can be rectified by approximating the best flow spanner. As the optimal connectivity between the nodes is ensured by interconnection networks, the fuzzy tree t- spanners of undirected fuzzy Butterfly Network BF(n) are examined by labelling the vertices and edges using ratio labelling. The Breadth First Search (BFS) algorithm enhances the process of minimizing the spanning tree for an unweighted graph. Since the edge weightage is unique for all the edges of BF(n) under RL, searching for an edge with minimum weight by Kruskal’s algorithm or Prim’s algorithm are less meaningful. Hence, the BFS algorithm is adopted in finding the spanning tree T of Ratio Labelled BF(n). The bounds for the stretch factor t of a fuzzy tree t- spanner of ratio labelled BF(n) were obtained. Novelty: The literature survey shows that various classes of graphs were examined for the admissibility of tree t-spanners. Finding a tree t-spanner for a fuzzy graph was not found in literature, as per our knowledge. Extending the problem of finding a tree t -spanner of a ratio labelled fuzzy graph is an advancement in the field of fuzzy graphs. Keywords: Fuzzy Graph, Fuzzy distance, Ratio Labelling, Tree t- spanner, Butterfly Network

Verifying hierarchical network nonlocality in general quantum networks

Abstract Recently, a class of innovative notions on quantum network nonlocality (QNN), called full quantum network nonlocality (FQNN), have been proposed in Phys. Rev. Lett. 128 010403 (2022). As the generalization of full network nonlocality (FNN), l-level quantum network nonlocality (l-QNN) was defined in arxiv. 2306.15717 quant-ph (2024). FQNN is a NN that can be generated only from a network with all sources being non-classical. This is beyond the existing standard network nonlocality, which may be generated from a network with only a non-classical source. One of the challenging tasks is to establish corresponding Bell-like inequalities to demonstrate the FQNN or l-QNN. Up to now, the inequality criteria for FQNN and l-QNN have only been established for star and chain networks. In this paper, we devote ourselves to establishing Bell-like inequalities for networks with more complex structures. Note that star and chain networks are special kinds of tree-shaped networks. We first establish the Bell-like inequalities for verifying l-QNN in k-forked tree-shaped networks. Such results generalize the existing inequalities for star and chain networks. Furthermore, we find the Bell-like inequality criteria for l-QNN for general acyclic and cyclic networks. Finally, we discuss the demonstration of l-QNN in the well-known butterfly networks.

Leveraging Model Scaling and Butterfly Network in the Bone Scan Image Segmentation

As we all know, cancer is one of the leading causes of death worldwide and the second leading cause of death overall. This is why regular screenings or health checks are necessary to detect cancer lesions early. Since bone scan images have become the primary means of detecting the emergence of cancer lesions on bone, high segmentation accuracy is essential for establishing the model of some predefined regions in bone scan images where cancer metastasis was predicted to appear. Consequently, robust localization and identification of the specific region in bone scan images are required for automated metastasis detection. To this end, we propose Efficient-BtrflyNet, a new deep learning-based architecture for skeleton segmentation of whole-body bone scan images. The proposed architecture exploits the benefits of EfficientNet’s model scaling and the encoder–decoder design of butterfly-type networks. We added EfficientNetB7 to the encoder section to obtain more specific features. The proposed architecture simultaneously processes anterior and posterior whole-body bone scan images. Using 37 bone scan images, we evaluated the performance of our proposed skeleton segmentation system using the Dice score. Efficient-BtrflyNet achieves superior segmentation performance compared to the existing representative method.

Antimagic labeling in butterfly and benes networks

This paper focus on the dynamic and rapidly evolving field of Graph Labeling research. In this paper, Antimagic Labeling for Benes and Butterfly networks has been determined. The findings can be utilized to determine the Antimagic Labeling in Interconnection networks where the routing servers acting as its vertices and the connection between them acts as edges to get the best possible solution for building a network with computational power and the ability to send and receive messages over cabled or wireless communication links.

Cops and Robber on butterflies, grids, and AT-free graphs

Cops and Robber is a well-studied two player pursuit-evasion game played on a graph. In this game, a set of cops, controlled by the first player, tries to capture the position of a robber, controlled by the second player. The cop number of a graph is the minimum number of cops required to capture the robber in the graph. A group of cops guard a subgraph if they can ensure that the robber cannot enter the subgraph without getting captured immediately.We study the applications of guarding to provide new bounds and improve existing bounds for several graph classes. In particular, we show that the cop number for butterfly networks and for solid grids is two. We also construct a partial grid with cop number 3 establishing that partial grids have same cop number as their superclass planar graphs. We also consider three well-studied variants of Cops and Robber: Cops and Fast Robber, Cops and Attacking Robber, and Surrounding CnR. We improve the existing bounds for the cop number of multidimensional grids for both Cops and Attacking Robber and Surrounding CnR. Finally, we consider Cops and Fast Robber on AT-free graphs to improve the existing bounds on the cop number for this game.

A training program for obstetrics point-of-care ultrasound to 514 rural healthcare providers in Kenya

BackgroundUltrasound is a crucial and effective diagnostic tool in medicine. Recent advancements in technology have led to increased use of point-of-care ultrasound (POCUS). Access to ultrasound equipment and training programs in low-and middle-income countries (LMICs) is limited. Despite the World Health Organization (WHO) recommendations for universal antenatal ultrasounds, POCUS for reproductive health applications has not been widely used in LMICs. We describe here the feasibility of implementation of a training of obstetrics point-of-care ultrasound (OB POCUS) for high-risk conditions in rural public healthcare facilities in Kenya with partnership from Butterfly Network, Global Ultrasound Institute, and Kenyatta University.MethodsAs part of the initiation of a large-scale implementation study of OB POCUS, clinician trainees were recruited from rural Kenyan hospitals for participation in a series of five-day POCUS workshops held between September and December 2022. Trainers provided brief didactic lessons followed by hands-on training with live models and at regional clinical sites for 5 OB POCUS applications. Instructor-observed assessment of students’ scanning and image interpretation occurred over the training period. Assessment of knowledge and confidence was performed via an online pre-test and post-test as well as Objective Structured Clinical Examination (OSCE) was administered at course completion.ResultsFive hundred and fourteen mid-level Health Care Providers (HCPs) in Kenya were trained over a three-month period through in-person didactic sessions, bedside instruction, and clinical practice over a 5-day period with a trainer: trainee ratio of approximately 1:5. Out of the 514 trained HCPs, 468 were from 8 rural counties with poor maternal and neonatal outcomes, while the remaining 46 were from nearby facilities. OB POCUS topics covered included: malpresentation, multiple gestation, fetal cardiac activity, abnormalities of the placenta and amniotic fluid volume. There was marked improvement in the post training test scores compared to the pretest scores.ConclusionOur implementation description serves as a guide for successful rapid dissemination of OB POCUS training for mid-level providers. Our experience demonstrates the feasibility of a short intensive POCUS training to rapidly establish specific POCUS skills in efforts to rapidly scale POCUS access and services. There is a widespread need for expanding access to ultrasound in pregnancy through accessible OB POCUS training programs. An implementation study is currently underway to assess the patient and systems-level impact of the training.

Effects of food availability on butterfly diversity and network specialization across altitudinal levels in a Mediterranean landscape

Effects of food availability on butterfly diversity and network specialization across altitudinal levels in a Mediterranean landscape

Poster 274: Preseason patellar tendon thickness as a predictor of symptomatic patellar tendinopathy in NCAA Division I Basketball players

Objectives: Patellar tendinopathy is prevalent in elite jumping athletes, with a particularly high risk in basketball players. Previous prospective research has demonstrated that ultrasound (US) findings, primarily the presence of hypoechoic regions within the tendon, are prevalent in both symptomatic and asymptomatic elite jumping athletes and thus do not necessarily indicate patellar tendon pathology. Other studies have suggested that hypoechoic regions within asymptomatic tendons are predictive of new onset symptomatic patellar tendinopathy. None of these previous studies evaluated the utility of patellar tendon thickness as a predictive measure. The purpose of this study is to build on the previous literature to identify whether preseason US is a clinically relevant preseason screening tool in elite basketball players. We hypothesized that increased preseason anteroposterior (AP) thickness of the proximal patellar tendon, as well as hypoechoic regions, are predictive of symptomatic patellar tendinopathy and associated sequelae. Methods: Thirty-one male National Collegiate Athletic Association (NCAA) Division I basketball players voluntarily participated in this prospective cohort study. Both patellar tendons of each athlete were evaluated at preseason, midseason, and postseason time points, resulting in a total of 52 patellar tendons from 27 athletes included in this analysis after exclusion of 4 tendons that were previously operated on and 6 tendons from 3 athletes who transferred to a new team during the season. Basic demographic data were collected for each athlete. At each time point, Victorian Institute of Sports Assessment – Patellar Tendon (VISA-P) scores, patellar tendon AP thickness, and presence of proximal patellar tendon hypoechoic regions were evaluated. Outcome measures included 1) proximal patellar tendon hypoechoic region, 2) trip to the training room (TTR), 3) time-loss symptomatic patellar tendinopathy (TLPT), and 4) patellar tendon rupture. TTR was defined as symptomatic patellar tendinopathy causing the athlete to seek evaluation by team medical staff. TLPT was a binary outcome that involved 1 or more missed game due to patellar tendinopathy symptoms. Measurement of patellar tendon AP thickness was performed using the Butterfly iQ portable US probe (Butterfly Network, Inc., Burlington, MA) and accompanying iPhone mobile application (Apple, Cupertino, CA) (Figure 1). Two distinct images were saved for each patellar tendon at each time point. Each of these images were independently evaluated by two reviewers, one of which was a fellowship- trained musculoskeletal radiologist. Intra- and interrater reliability was calculated. Covariates evaluated in multivariate regression model included age and body mass index (BMI). Significance was set to p<.05. Results: Of the total of 52 tendons (mean age 20.67 ± 1.22 years, mean BMI 24.1 ± 1.57), the mean preseason AP thickness was 4.78 ± 1.22 mm. Intra- and interrater reliability for AP thickness were both excellent (ICC=0.99, p<.01 and ICC=0.98, p<.01, respectively). Compared to preseason values, there was a significant increase in mean AP thickness at both midseason (p<.01) and postseason (p<.01) evaluation (Table 1, Figure 2). VISA-P scores decreased from preseason to midseason (p=.04), but increased from midseason to postseason (p=.03) (Table 1). A total of 9 patellar tendons resulted in a TTR (17.31%) (Table 2). Preseason AP thickness was associated with an increased odds of TTR (aOR=7.58 [2.09, 29.66], p<.01). Assuming mean age and BMI, an 8mm tendon at baseline predicted a 97.19% probability of a TTR during the season, while a 4mm tendon predicted a 0.90% probability. Of the 9 tendons that demonstrated preseason hypoechoic regions, 5 (55.56%) resulted in a TTR (aOR=12.60, p<.01). Preseason VISA-P scores (p=.36) were not predictive of TTR. There were 4 cases of TLPT (7.69%) (Table 2). There was a significant association between preseason AP thickness and increased odds of TLPT (aOR=1.61 [1.02, 2.54], p=.04). Assuming mean age and BMI, an 8mm tendon at baseline demonstrated a 24.83% probability of resulting in TLPT, while a 4mm tendon demonstrated a 4.26% probability. Preseason VISA-P scores (p=.39) were not predictive of TLPT. Conclusions: This is the first analysis evaluating patellar tendon thickness and patellar tendon sequelae in elite basketball players at more than two time points. This study demonstrated the ability for preseason US screening to identify symptomatic patellar tendinopathy and associated sequelae during the season. Additionally, there was a significant increase in mean AP thickness from preseason to midseason as well as from midseason to postseason. This is in contrast to VISA-P scores, which initially increased, but then decreased to preseason levels at postseason evaluation. Patellar tendon AP thickness may thus serve as a more stable and precise measure of patellar tendinopathy severity when compared to VISA-P scores. Further research with increased sample size may establish US as a clinically relevant screening tool in elite jumping athletes. [Table: see text][Table: see text]

Noninvasive Hemodynamic Profiling of Patients Undergoing Hemodialysis Using a Handheld Ultrasound Device

INTRODUCTION: Accurate determination of volume status for patients with end-stage renal disease is essential in determining ultrafiltration rate during hemodialysis (HD). To complement the current dry weight method, inferior vena cava (IVC) collapsibility, made accessible by point of-care ultrasonography, is considered. This study determined the utility of IVC measurement in estimating the volume status of patients during HD in comparison to clinical parameters. METHODS: A single-center cross-sectional design including 53 HD patients was conducted, with IVC measurements done through the Butterfly iQ ultrasound (Butterfly Network, Burlington, Massachusetts). RESULTS: Most patients were hypervolemic before HD based on weight (94.3%) and IVC collapsibility index (IVC-CI; 75.5%), but only 30% had clinical symptoms. Body weight, maximum IVC diameter, minimum IVC diameter, and indexed IVC size significantly decreased after HD, whereas IVC-CI, blood pressure, and heart rate were unchanged. For the subset of patients with symptoms, absolute values of IVC measures were higher, but did not significantly change after HD, unlike in those without symptoms. For volume classification, there are discrepancies in the classifications based on the different measures, with most improvement seen when weight was used, but which was not reflected in IVC-CI. Change in weight and IVC measures were not significantly correlated. DISCUSSION: This pilot study showed that the current dry weight method provides ultrafiltration rate estimation without causing intradialytic events. However, IVC can be a supplemental parameter to set higher targets and increase volume removal enough to cause intravascular change, especially in symptomatic patients. The incongruencies in classifying volume status suggest that there is no single measure to determine hemodynamic status and that using multiple parameters may provide a more reliable estimate. KEYWORDS: inferior vena cava measurement, volume status in hemodialysis, point-of-care ultrasonography

A Novel Fuzzy DBNet for Medical Image Segmentation

When doctors are fatigued, they often make diagnostic errors. Similarly, pharmacists may also make mistakes in dispensing medication. Therefore, object segmentation plays a vital role in many healthcare-related areas, such as symptom analysis in biomedical imaging and drug classification. However, many traditional deep-learning algorithms use a single view of an image for segmentation or classification. When the image is blurry or incomplete, these algorithms fail to segment the pathological area or the shape of the drugs accurately, which can then affect subsequent treatment plans. Consequently, we propose the Fuzzy DBNet, which combines the dual butterfly network and the fuzzy ASPP in a deep-learning network and processes images from both sides of an object simultaneously. Our experiments used multi-category pill and lung X-ray datasets for training. The average Dice coefficient of our proposed model reached 95.05% in multi-pill segmentation and 97.05% in lung segmentation. The results showed that our proposed model outperformed other state-of-the-art networks in both applications, demonstrating that our model can use multiple views of an image to obtain image segmentation or identification.

The Integration of Point-of-Care Ultrasound into Practical Trauma Training in Ukraine: A Case Study Highlighting Feasibility, Satisfaction, and Pre- and Post-Training Comfort with the Technology

Introduction:As of October 2022, the civilian casualty count of the invasion of Ukraine is reported to be 16,295, with actual figures believed to be considerably higher. As explosive trauma continues to terrorize populations, frontline medical personnel are faced with escalating resource constraints including transport, imaging modalities, and electricity. Point of care ultrasound (POCUS) is considered the gold standard in acute trauma evaluation, but very few hospitals or pre-hospital medics have access to or training in POCUS.Method:In collaboration with the Ukrainian Ministry of Health, the World Health Organization, and the Global Health Program at Butterfly Network, Team Rubicon developed and conducted 64 practical trauma trainings and donated 50 Butterfly iQ+ portable ultrasound devices in Ukraine between August and October, 2022. Of these trainings, 19 specifically focused on the use of POCUS for trauma. Pre- and post-surveys were deployed to determine demographics, comfort level with POCUS for trauma care, and usefulness of the course.Results:In total, 149 individuals were trained in POCUS for trauma. Of these, 130 were physicians, 15 were paramedics, three were RNs, and one was a pharmacist. Only 14.8% of these clinicians self-reported any previous POCUS training. All participants reported an increase in comfort level, with an average pre- and post-course comfort scores of 1.9 and 3.3 (4=very comfortable), respectively. General satisfaction with the training was high (average score of 9.8/10). Qualitative feedback commended the quality and novelty of this training, requested further examples of pathology, and endorsed more POCUS trainings, generally. The most critical lesson learned was the need to re-orient training around the foundations of POCUS given low levels of experience and training.Conclusion:Access and training in POCUS for trauma is critical for resource-constrained medical personnel operating in conflict-affected communities. A one-day POCUS practicum-oriented course is feasible to support awareness and proficiency.

Butterfly Transforms for Efficient Representation of Spatially Variant Point Spread Functions in Bayesian Imaging

Bayesian imaging algorithms are becoming increasingly important in, e.g., astronomy, medicine and biology. Given that many of these algorithms compute iterative solutions to high-dimensional inverse problems, the efficiency and accuracy of the instrument response representation are of high importance for the imaging process. For efficiency reasons, point spread functions, which make up a large fraction of the response functions of telescopes and microscopes, are usually assumed to be spatially invariant in a given field of view and can thus be represented by a convolution. For many instruments, this assumption does not hold and degrades the accuracy of the instrument representation. Here, we discuss the application of butterfly transforms, which are linear neural network structures whose sizes scale sub-quadratically with the number of data points. Butterfly transforms are efficient by design, since they are inspired by the structure of the Cooley-Tukey fast Fourier transform. In this work, we combine them in several ways into butterfly networks, compare the different architectures with respect to their performance and identify a representation that is suitable for the efficient representation of a synthetic spatially variant point spread function up to a 1% error. Furthermore, we show its application in a short synthetic example.

Construct Validity and Reproducibility of Handheld Ultrasound Devices in Carotid Artery Diameter Measurement.

The construct validity and reproducibility of three commonly used handheld ultrasound (US) devices in measuring carotid arterial diameter was evaluated: Telemed MicrUs EXT-1H (Telemed, Vilnius, Lithuania), Butterfly iQ (Butterfly Network, Inc., Guilford, CT, USA) and Philips Lumify (Philips Healthcare, Best, The Netherlands). An in vitro setup was built to evaluate construct validity, compared with high-end US, and intra-observer variability of handheld US devices. Handheld devices showed a mean difference of 0.023 ± 0.030 cm, 0.012 ± 0.037 cm and 0.009 ± 0.046 cm for, respectively, Telemed, Butterfly and Lumify in comparison with high-end US devices. Intraclass agreement with the high-end system as well as intra-observer variability for handheld US devices was classified as excellent, with all values greater than 0.95. Subsequently, inter-observer variability of handheld US devices was investigated in an in vivo setup with 20 healthy volunteers. Inter-observer variability was classified as excellent for Telemed (0.901), good for Lumify (0.827) and moderate for Butterfly (0.684) with a difference of, respectively, 0.005 ± 0.031 cm, 0.020 ± 0.050 cm and -0.003 ± 0.033 cm. In conclusion, handheld US devices demonstrated an excellent construct validity and intra-observer variability. Additionally, excellent-to-good inter-observer variability for Telemed and Lumify was observed, and Butterfly demonstrated a moderate inter-observer agreement. These results indicate that handheld US devices are effective for measuring carotid arterial diameter.

Scale-out Systolic Arrays

Multi-pod systolic arrays are emerging as the architecture of choice in DNN inference accelerators. Despite their potential, designing multi-pod systolic arrays to maximize effective throughput/Watt—i.e., throughput/Watt adjusted when accounting for array utilization—poses a unique set of challenges. In this work, we study three key pillars in multi-pod systolic array designs, namely array granularity, interconnect, and tiling. We identify optimal array granularity across workloads and show that state-of-the-art commercial accelerators use suboptimal array sizes for single-tenancy workloads. We, then evaluate the bandwidth/latency trade-offs in interconnects and show that Butterfly networks offer a scalable topology for accelerators with a large number of pods. Finally, we introduce a novel data tiling scheme with custom partition size to maximize utilization in optimally sized pods. We propose Scale-out Systolic Arrays , a multi-pod inference accelerator for both single- and multi-tenancy based on these three pillars. We show that SOSA exhibits scaling of up to 600 TeraOps/s in effective throughput for state-of-the-art DNN inference workloads, and outperforms state-of-the-art multi-pod accelerators by a factor of 1.5 ×. 1

Butterfly network: a convolutional neural network with a new architecture for multi-scale semantic segmentation of pedestrians.

The detection of multi-scale pedestrians is one of the challenging tasks in pedestrian detection applications. Moreover, the task of small-scale pedestrian detection, i.e., accurate localization of pedestrians as low-scale target objects, can help solve the issue of occluded pedestrian detection as well. In this paper, we present a fully convolutional neural network with a new architecture and an innovative, fully detailed supervision for semantic segmentation of pedestrians. The proposed network has been named butterfly network (BF-Net) because of its architecture analogous to a butterfly. The proposed BF-Net preserves the ability of simplicity so that it can process static images with a real-time image processing rate. The sub-path blocks embedded in the architecture of the proposed BF-Net provides a higher accuracy for detecting multi-scale objective targets including the small ones. The other advantage of the proposed architecture is replacing common batch normalization with conditional one. In conclusion, the experimental results of the proposed method demonstrate that the proposed network outperform the other state-of-the-art networks such as U-Net + + , U-Net3 + , Mask-RCNN, and Deeplabv3 + for the semantic segmentation of the pedestrians.