Population data-based federated machine learning improves automated echocardiographic quantification of cardiac structure and function: the Automatisierte Vermessung der Echokardiographie project

Aims

Machine-learning (ML)-based automated measurement of echocardiography images emerges as an option to reduce observer variability. The objective of the study is to improve the accuracy of a pre-existing automated reading tool ('original detector') by federated ML-based re-training.

Methods and results

Automatisierte Vermessung der Echokardiographie was based on the echocardiography images of n = 4965 participants of the population-based Characteristics and Course of Heart Failure Stages A-B and Determinants of Progression Cohort Study. We implemented federated ML: echocardiography images were read by the Academic Core Lab Ultrasound-based Cardiovascular Imaging at the University Hospital Würzburg (UKW). A random algorithm selected 3226 participants for re-training of the original detector. According to data protection rules, the generation of ground truth and ML training cycles took place within the UKW network. Only non-personal training weights were exchanged with the external cooperation partner for the refinement of ML algorithms. Both the original detectors as the re-trained detector were then applied to the echocardiograms of n = 563 participants not used for training. With regard to the human referent, the re-trained detector revealed (i) superior accuracy when contrasted with the original detector's performance as it arrived at significantly smaller mean differences in all but one parameter, and a (ii) smaller absolute difference between measurements when compared with a group of different human observers.

Conclusion

Population data-based ML in a federated ML set-up was feasible. The re-trained detector exhibited a much lower measurement variability than human readers. This gain in accuracy and precision strengthens the confidence in automated echocardiographic readings, which carries large potential for applications in various settings.

Automated measurement of the foveal avascular zone in healthy eyes on Heidelberg spectralis optical coherence tomography angiography

PURPOSE

To develop and evaluate an automated method to measure the foveal avascular zone (FAZ) area in healthy eyes on Heidelberg Spectralis Optical Coherence Tomography Angiography (HS-OCTA). This method is referred to as the modified Kanno-Saitama macro (mKSM) which is an evolution of the Kanno-Saitama macro (KSM) approach.

METHODS

This cross-sectional study included 29 eyes of 25 healthy volunteers who underwent HS-OCTA at the macular area twice at the same time. Regardless of the quality of the images, all of them were included. Macular data on the superficial vascular plexus, intermediate capillary plexus (ICP) and deep capillary plexus were processed by mKSM. The FAZ area was measured twice automatically using the mKSM and KSM and twice manually by two independent examiners.

RESULTS

From 174 images, KSM could not measure correctly 31% while mKSM could successfully measure all of them. Intrascan intraclass coefficient ranged from 0,948 to 0,993 for manual measurements and was 1 for mKSM method. Despite that the difference between human examiners is smaller than between human examiners and mKSM according to Bland-Altman plots, the scatterplots show a strong correlation between human and automatic measurements. The best results are obtained in ICP.

CONCLUSIONS

With mKSM, the automated determination of the FAZ area in HS-OCTA is feasible and less human-dependent. It solves the inability of KSM to measure the FAZ area in suboptimal quality images which are frequent in daily clinical practice. Therefore, the mKSM processing could contribute to our understanding of the three vascular plexuses.

Radiolucent Zone around Screws is Associated with Position Change of Screw-rod Constructs

PURPOSE

A radiologic assessment method to measure position change of screw-rod constructs over time by superposing the 3‑dimensional images assists in quantitative evaluation of screw loosening. We investigated the association between position change and radiolucent zone that was commonly used for diagnosing screw loosening.

METHODS

In this study 101 patients who underwent lumbar fusion were reviewed. Patient characteristics included age, sex, indications for surgery, number of fused levels, surgical procedures, and timing of follow-up computed tomography (CT, 1-5 months, 6-11 months, and ≥ 12 months). The Hounsfield unit values of L1 vertebra on preoperative CT were measured, and the radiolucent zone on each follow-up CT was evaluated. Using baseline CT on the day after surgery and follow-up CT, 3‑dimensional images of screw-rod constructs were generated and superposed. Position change was assessed by the median of the distances between the 3‑dimensional images at baseline and follow-up using the automated measurement method. Patient characteristics, the Hounsfield unit values of L1, and the amount of position change were categorized into the radiolucent zone presence and absence groups and compared.

RESULTS

The medians of position change were 0.281 mm and 0.136 mm in the radiolucent zone presence and absence groups, respectively (P < 0.001 by Mann-Whitney U-test). The area under the curve for position change in identifying radiolucent zone was 0.846; the cut-off value was 1.76 mm. In multivariable analysis, position change was independently associated with radiolucent zone (adjusted odds ratio per 0.1 mm, 2.80, 95% confidence interval 1.70-4.61).

CONCLUSION

Radiolucent zone was associated with position change of screw-rod constructs.

Computational approaches to understanding interaction and development

Audio-visual recording and location tracking produce enormous quantities of digital data with which researchers can document children's everyday interactions in naturalistic settings and assessment contexts. Machine learning and other computational approaches can produce replicable, automated measurements of these big behavioral data. The economies of scale afforded by repeated automated measurements offer a potent approach to investigating linkages between real-time behavior and developmental change. In our work, automated measurement of audio from child-worn recorders-which quantify the frequency of child and adult speech and index its phonemic complexity-are paired with ultrawide radio tracking of children's location and interpersonal orientation. Applications of objective measurement indicate the influence of adult behavior in both expert ratings of attachment behavior and ratings of autism severity, suggesting the role of dyadic factors in these "child" assessments. In the preschool classroom, location/orientation measures provide data-driven measures of children's social contact, fertile ground for vocal interactions. Both the velocity of children's movement toward one another and their social contact with one another evidence homophily: children with autism spectrum disorder, other developmental disabilities, and typically developing children were more likely to interact with children in the same group even in inclusive preschool classrooms designed to promote interchange between all children. In the vocal domain, the frequency of peer speech and the phonemic complexity of teacher speech predict the frequency and phonemic complexity of children's own speech over multiple timescales. Moreover, children's own speech predicts their assessed language abilities across disability groups, suggesting how everyday interactions facilitate development.

Measuring sociability of mice using a novel three-chamber apparatus and algorithm of the LABORAS™ system

BACKGROUND

The social approach and social novelty tests utilizing the three-chamber apparatus are widely accepted to measure social behavior of rodents. The LABORAS™ system offers a possibility to assess sociability of mice in a reliable and objective manner.

NEW METHOD

We assessed the capability of the LABORAS™ sociability cage and algorithm (2.6.6) to detect social behaviors in mice. Furthermore, we investigated whether the system is able to detect various levels of sociability due to genetic background or after pharmacological treatments.

RESULTS

By comparing manual scoring with various detection zone settings of the automated registration, the most fitting algorithm with a detection zone radius of 90 mm was identified. When different strains were investigated, C57Bl/6 J and NMRI mice proved to be social, while CD1 mice were found asocial. The system was able to detect the sociability increasing effect of R-baclofen (0.5 mg/kg i.p.) and oxytocin (12 ng i.c.v.) in asocial CD1 mice. The negative control PCP impaired social behavior of C57Bl/6 J mice (1 mg/kg i.p.) and increased social avoidance in CD1 mice (0.3 mg/kg i.p.).

COMPARISON WITH EXISTING METHOD(S)

This setup, in contrast to video frame analysis softwares, determines signal changes caused by movements of rodents allowing accurate detection and analysis of trajectories. Parallel automated measurements also allow replacing time and labor intensive, highly subjective human observational work.

CONCLUSIONS

The set-up provides a fast and reliable method to examine social behavior of mice in the three-chamber apparatus. The system is capable of detecting pro or antisocial activity of pharmacological agents.

Design and clinical validation of a software program for automated measurement of mammographic breast density

Background

Mammographic breast density is an important predictor of breast cancer, but its measurement has limitations related to subjectivity of visual evaluation or to difficult access for automatic volumetric measurement methods. Herein, we describe the design and clinical validation of Aguida, a software program for automated quantification of breast density from flat mammography images.

Materials and methods

The software program was developed in MatLab. After image segmentation separating the background from the breast image, the operator positions a cursor defining a region of interest on the pectoralis major muscle from the mediolateral oblique view. Then, in the craniocaudal view, the threshold for separation of the dense tissue is based on the optical density of the pectoral muscle, and the proportion of dense tissue is calculated by the program. Mammograms obtained from 2 different occasions in 291 women were used for clinical evaluation.

Results

The intraclass correlation coefficient (ICC) between breast density measurements by the software and by a radiologist was 0.96, with a bias of only 0.67 percentage points and a 95% limit of agreement of 13.5 percentage points; the ICC was 0.94 in the interobserver reliability assessment by two radiologists with different experience; and the ICC was 0.98 in the intraobserver reliability assessment. The distribution among the density classes was close to the values obtained with the volumetric software.

Conclusions

Measurement of breast density with the Aguida program from flat mammography images showed high agreement with the visual determination by radiologists, and high inter- and intra-observer reliability.

Aortic roots assessment by an automated three-dimensional transesophageal echocardiography: an intra-individual comparison

To evaluate the accuracy, reproducibility, and transcatheter heart valve (THV) sizing efficiency of an automated 3-dimensional transesophageal echocardiographic (3D-TEE) post-processing software in the assessments of aortic roots, intra-individually compared with multidetector computed tomography (MDCT). We prospectively studied 67 patients with normal aortic roots. We measured diameters of aortic annulus (AA), sinus of Valsalva (SOV), and sino-tubular junction (STJ) by full-automated and semi-automated methods using 3D-TEE datasets, then compared them to corresponding transthoracic echocardiography and MDCT values. THV sizes were chosen based on echocardiography and MDCT measurements according to recommended criterion. Taking MDCT planimetered diameters as reference, the full-automated (r: 0.4745-0.8792) and semi-automated (r: 0.6647-0.8805) 3D-TEE measurements were linearly correlated (p < 0.0001). The average differences between semi-automated or full-automated measurements and reference were 0.3 mm or 1.3 mm for AA, - 1.9 mm or - 0.5 mm for SOV, and - 0.1 mm or 1.9 mm for STJ, respectively. The intra-class correlation coefficients of semi-automated method were 0.79-0.96 (intra-observer) and 0.75-0.92 (inter-observer). THV sizing by semi-automated measurements using echocardiographic criteria was larger than that by MDCT measurements using MDCT criteria (p < 0.0001) but equivalent (p > 0.05) if both using MDCT standards. The new automated 3D-TEE software allows modeling and quantifying aortic roots with high reproducibility. Measurements by the semi-automated method closely approximate and well correlate with the corresponding MDCT, thus THV sizing by this modeled 3D-TEE measurements should adopt recommended MDCT criteria but not echocardiographic criteria. The full-automated 3D-TEE segmentations are yet immature. (Semi-automated assessMent of Aortic Roots by Three-dimensional transEsophageal echocaRdiography [SMARTER], NCT02724709).

An instrumented centrifuge for studying mouse locomotion and behaviour under hypergravity

Gravity may influence multiple aspects of legged locomotion, from the periods of limbs moving as pendulums to the muscle forces required to support the body. We present a system for exposing mice to hypergravity using a centrifuge and studying their locomotion and activity during exposure. Centrifuge-induced hypergravity has the advantages that it both allows animals to move freely, and it affects both body and limbs. The centrifuge can impose two levels of hypergravity concurrently, using two sets of arms of different lengths, each carrying a mouse cage outfitted with a force and speed measuring exercise wheel and an infrared high-speed camera; both triggered automatically when a mouse begins running on the wheel. Welfare is monitored using infrared cameras. As well as detailing the design of the centrifuge and instrumentation, we present example data from mice exposed to multiple levels of hypergravity and details of how they acclimatized to hypergravity.

The variability of automated QRS duration measurement

Aims

Previous studies have demonstrated substantial variability in manual assessment of QRS complex duration (QRSd). Disagreements in QRSd measurements were also found in several automated algorithms tested on digitized electrocardiogram (ECG) recordings. The aim of our study was to investigate the variability of automated QRSd measurements performed by two commercially available electrocardiographs.

Methods and Results

Two GE MAC 5000 (GE-1 and GE-2) electrocardiographs and two Mortara ELI 350 (Mortara-1 and Mortara-2) electrocardiographs were used in the study. Participants for the study were recruited from patients hospitalized in the department of cardiology of a university hospital. Participants underwent up to four recording sessions within a single day with a different electrocardiograph at each session when two to four immediately successive ECG recordings were undertaken. In 76 patients, 683 ECGs were recorded; the mean QRSd was 109.0 ± 26.1 ms. The QRSd difference ≥10 ms between the first and second intra-session ECG was found in 7, 3, 20, and 14% of ECG pairs for GE-1, GE-2, Mortara-1, and Mortara-2, respectively. No inter-session difference in QRSd was found within both manufacturers. In individual patients, Mortara calculated the mean QRSd to be longer by 7.3 ms (95% CI: 6.2-8.5 ms, P < 0.0001) with a 2.1-times (95% CI: 1.9-2.4) greater standard deviation of the mean QRSd (7.1 vs. 3.3 ms, P < 0.001).

Conclusion

Electrocardiographs from two manufacturers measured QRSd values with a systematic difference and a significantly different level of precision. This may have important clinical implications in selection of suitable candidates for cardiac resynchronization therapy.

Implementing an electronic hand hygiene monitoring system: Lessons learned from community hospitals

BACKGROUND

Measuring and providing feedback about hand hygiene (HH) compliance is a complicated process. Electronic HH monitoring systems have been proposed as a possible solution; however, there is little information available about how to successfully implement and maintain these systems for maximum benefit in community hospitals.

METHODS

An electronic HH monitoring system was implemented in 3 community hospitals by teams at each facility with support from the system vendor. Compliance rates were measured by the electronic monitoring system. The implementation challenges, solutions, and drivers of success were monitored within each facility.

RESULTS

The electronic HH monitoring systems tracked on average more than 220,000 compliant HH events per facility per month, with an average monthly compliance rate >85%. The sharing of best practices between facilities was valuable in addressing challenges encountered during implementation and maintaining a high rate of use.

DISCUSSION

Drivers of success included a collaborative environment, leadership commitment, using data to drive improvement, consistent and constant messaging, staff empowerment, and patient involvement.

CONCLUSIONS

Realizing the full benefit of investments in electronic HH monitoring systems requires careful consideration of implementation strategies, planning for ongoing support and maintenance, and presenting data in a meaningful way to empower and inspire staff.

Fully Automated Deep Learning System for Bone Age Assessment

Skeletal maturity progresses through discrete phases, a fact that is used routinely in pediatrics where bone age assessments (BAAs) are compared to chronological age in the evaluation of endocrine and metabolic disorders. While central to many disease evaluations, little has changed to improve the tedious process since its introduction in 1950. In this study, we propose a fully automated deep learning pipeline to segment a region of interest, standardize and preprocess input radiographs, and perform BAA. Our models use an ImageNet pretrained, fine-tuned convolutional neural network (CNN) to achieve 57.32 and 61.40% accuracies for the female and male cohorts on our held-out test images. Female test radiographs were assigned a BAA within 1 year 90.39% and within 2 years 98.11% of the time. Male test radiographs were assigned 94.18% within 1 year and 99.00% within 2 years. Using the input occlusion method, attention maps were created which reveal what features the trained model uses to perform BAA. These correspond to what human experts look at when manually performing BAA. Finally, the fully automated BAA system was deployed in the clinical environment as a decision supporting system for more accurate and efficient BAAs at much faster interpretation time (<2 s) than the conventional method.

Neurological software tool for reliable atrophy measurement (NeuroSTREAM) of the lateral ventricles on clinical-quality T2-FLAIR MRI scans in multiple sclerosis

Background

There is a need for a brain volume measure applicable to the clinical routine scans. Nearly every multiple sclerosis (MS) protocol includes low-resolution 2D T2-FLAIR imaging.

Objectives

To develop and validate cross-sectional and longitudinal brain atrophy measures on clinical-quality T2-FLAIR images in MS patients.

Methods

A real-world dataset from 109 MS patients from 62 MRI scanners was used to develop a lateral ventricular volume (LVV) algorithm with a longitudinal Jacobian-based extension, called NeuroSTREAM. Gold-standard LVV was calculated on high-resolution T1 1 mm, while NeuroSTREAM LVV was obtained on low-resolution T2-FLAIR 3 mm thick images. Scan-rescan reliability was assessed in 5 subjects. The variability of LVV measurement at different field strengths was tested in 76 healthy controls and 125 MS patients who obtained both 1.5T and 3T scans in 72 hours. Clinical validation of algorithm was performed in 176 MS patients who obtained serial yearly MRI 1.5T scans for 10 years.

Results

Correlation between gold-standard high-resolution T1 LVV and low-resolution T2-FLAIR LVV was r = 0.99, p < 0.001 and the scan-rescan coefficient of variation was 0.84%. Correlation between low-resolution T2-FLAIR LVV on 1.5T and 3T was r = 0.99, p < 0.001 and the scan-rescan coefficient of variation was 2.69% cross-sectionally and 2.08% via Jacobian integration. NeuroSTREAM showed comparable effect size (d = 0.39–0.71) in separating MS patients with and without confirmed disability progression, compared to SIENA and VIENA.

Conclusions

Brain atrophy measurement on clinical quality T2-FLAIR scans is feasible, accurate, reliable, and relates to clinical outcomes.

•

A robust algorithm for measuring lateral ventricular volume on clinical FLAIR scans is proposed.

•

The algorithm combines multi-atlas joint fusion labeling with level-set smoothness-constraining refinement.

•

Results show a similar relationship to disability progression as with established metrics on high-resolution scans.

Open-Source Radiation Exposure Extraction Engine (RE3) with Patient-Specific Outlier Detection

We present an open-source, picture archiving and communication system (PACS)-integrated radiation exposure extraction engine (RE3) that provides study-, series-, and slice-specific data for automated monitoring of computed tomography (CT) radiation exposure. RE3 was built using open-source components and seamlessly integrates with the PACS. RE3 calculations of dose length product (DLP) from the Digital imaging and communications in medicine (DICOM) headers showed high agreement (R (2) = 0.99) with the vendor dose pages. For study-specific outlier detection, RE3 constructs robust, automatically updating multivariable regression models to predict DLP in the context of patient gender and age, scan length, water-equivalent diameter (D w), and scanned body volume (SBV). As proof of concept, the model was trained on 811 CT chest, abdomen + pelvis (CAP) exams and 29 outliers were detected. The continuous variables used in the outlier detection model were scan length (R (2)  = 0.45), D w (R (2) = 0.70), SBV (R (2) = 0.80), and age (R (2) = 0.01). The categorical variables were gender (male average 1182.7 ± 26.3 and female 1047.1 ± 26.9 mGy cm) and pediatric status (pediatric average 710.7 ± 73.6 mGy cm and adult 1134.5 ± 19.3 mGy cm).

A microchip for quantitative analysis of CNS axon growth under localized biomolecular treatments

Growth capability of neurons is an essential factor in axon regeneration. To better understand how microenvironments influence axon growth, methods that allow spatial control of cellular microenvironments and easy quantification of axon growth are critically needed. Here, we present a microchip capable of physically guiding the growth directions of axons while providing physical and fluidic isolation from neuronal somata/dendrites that enables localized biomolecular treatments and linear axon growth. The microchip allows axons to grow in straight lines inside the axon compartments even after the isolation; therefore, significantly facilitating the axon length quantification process. We further developed an image processing algorithm that automatically quantifies axon growth. The effect of localized extracellular matrix components and brain-derived neurotropic factor treatments on axon growth was investigated. Results show that biomolecules may have substantially different effects on axon growth depending on where they act. For example, while chondroitin sulfate proteoglycan causes axon retraction when added to the axons, it promotes axon growth when applied to the somata. The newly developed microchip overcomes limitations of conventional axon growth research methods that lack localized control of biomolecular environments and are often performed at a significantly lower cell density for only a short period of time due to difficulty in monitoring of axonal growth. This microchip may serve as a powerful tool for investigating factors that promote axon growth and regeneration.