How to evaluate (and perhaps how to add novelties of) AI-enabled tools for echocardiography data?

[mxochicale]

🚀 Feature

I am opening this issue to track and to discuss potential avenues for the evaluation of AI-enabled tools for echocardiography data.

Motivation

In recent meetings, Luigi has mentioned the new AI-based tools for venue go machines where the way GE evaluate their performance is by reporting the percentage in reduction of keystrokes and the percentage of less time compared to the manual method calculation. In the above site, GE cite the work of Bobbia, L Muller, PG Claret, L Vigouroux in Shock, 2019 that states that with the use of AI, "the ultrasound device optimized the pulsed Doppler box place, recorded four seconds of Doppler spectrum, traced the outline of VTIs, averaged all VTIs of the recording (VTIauto), and calculated the HR and CO (COauto)".

Another interesting line of discussion, from Bobbia et al in 2019, is to have understanding of conventional vs automatic echocardiography procedures.

Pitch

Identify an approach to evaluate the AI-enabled tools for echocardiography data that can help clinicians and also add novelty to the state-of-the-art of US community.

Alternatives

Not at the moment

Additional context

Not at the moment

[mxochicale]

Cardiac motion analysis based on 2D clips

Auto View Recognition: AutoStrain recognizes A4C, A2C and A3C view and assigns labels Auto Contour Placement: Automated contour definition and speckle tracking for user review Cardiac cycle can be defined manually in case of missing or bad ECG signal Support of TTE and TEE data

https://www.tomtec.de/products/application-finder/autostrain-lv/rv/la

Analysis of the left ventricle with three clicks — integrated in your clinical workflow

Standard LV volume measurements with minimal user interaction Automated contouring in end-diastole and end-systole Verification of automated results by manual contour editing possible Volumes and EF are displayed immediately

Precise results within seconds Guideline conform measurements Automated biplane Simpson’s method High reproducibility due to automation without user interaction Approved algorithm by multicenter study

Remain in your regular clinical workflow Stay in your regular review application No additional software has to be opened Directly access your results in the worksheet and report

https://www.tomtec.de/products/application-finder/autolv

[mxochicale]

AutoLV: Functional and anatomical analysis of the left ventricle in one click using the power of Artificial Intelligence!

AutoLV Analysis software makes functional and anatomical analysis of the left ventricle: up to 46x faster1 highly reproducible operator-independent

LV Trace in B-Mode: https://www.youtube.com/watch?v=i89WXc8uA9U&t=12s

LV Trace in M-Mode: https://www.youtube.com/watch?v=Svt0ljyh3AA&t=35s

1. Grune, J. et al. Accurate assessment of LV function using the first automated 2D-border detection algorithm for small animals - evaluation and application to models of LV dysfunction. Cardiovasc. Ultrasound 17, 7 (2019).

2. Lang RM, Bierig M, Devereux RB, et al. Recommendations for chamber quantification: A report from the American Society of Echocardiography’s guidelines and standards committee and the Chamber Quantification Writing Group, developed in conjunction with the European Association of Echocardiograph. J Am Soc Echocardiogr. 2005;18(12):1440-1463. doi:10.1016/j.echo.2005.10.005.

https://www.visualsonics.com/autolv

[mxochicale]

Philips

DHM: Dynamic HeartModelA.I. "DHM SB 11 30 2018"

https://www.youtube.com/watch?v=t9j1K15A6as https://www.documents.philips.com/doclib/enc/17271434/Dynamic-heartmodel-White-Paper.pdf

Automated transthoracic three-dimensional echocardiographic quantification of the left heart chambers "QLAB and with the new HeartModel"

Philips has recently developed HeartModel A.I. , a fully automated 3D-TTE analysis software which simultaneously detects LA and LV endocardial surfaces throughout the cardiac cycle, using an adaptive analytics algorithm that consists of knowledge-based identification of initial global shape and orientation followed by patient specific adaptation (Figure 1). The process begins with the program estimating the LV end-diastolic (ED) frame using motion analysis near the peak of the electrocardiographic R-wave. Using this frame, general shape orientation is identified, and then the LV end-systolic (ES) frame is estimated using motion analysis to identify the smallest LV cavity. Once LV ED and ES frames have been estimated, preliminary ES and ED models of the LV and LA are built using the automatically detected endocardial surface in conjunction with information from a 3D TTE LA and LV database. This database consists of a variety of LA and LV ED and ES shapes obtained from approximately 1000 3D-TTE data of varying image quality in patients with a wide range of chamber size and function. The program matches features from the LV volume being analyzed to selected shapes in the database. This selected shape is then locally adapted to the LV volume under study using a series of incremental steps. The ED and ES frames are then finally detected by evaluating the LV volumes in the neighborhood of end- diastole and end-systole and selecting the frame with the maximum and minimum volumes, respectively. The algorithm was designed to adjust to a variety of imaging conditions, including variations in dropout, acoustic clutter, ventricular shape, and cardiac orientation relative to the transducer. However, similar to manual measurements, a minimum number of visible endocardial border segments (∼ 14-15 of 17 LV segments) is necessary for an accurate estimate of the chamber volume to be derived. Lastly, when run on the same data set, the algorithm has a deterministic convergence response, thus yielding zero variability. Once the final model has been fitted, the LA and LV contours are displayed on 2D cut-planes derived from the 3DE data sets showing the ES and ED 4-, 3- and 2-chamber views (Figure 2). If the user is not satisfied with the LA and LV contours, they could be manually edited.

© 2016 Koninklijke Philips N.V. All rights are reserved. Philips reserves the right to make changes in specifications and/or to discontinue any product at any time without notice or obligation and will not be liable for any consequences resulting from the use of this publication. Trademarks are the property of Koninklijke Philips N.V. or their respective owners. Roberto M. Lang 5841 South Maryland Ave, MC5084 Chicago, Illinois, USA 60611 Email: rlang@medicine.bsd.uchicago.edu Telephone: 773-702-1842 Fax: 773-773-1084

https://www.documents.philips.com/doclib/enc/12475792/452299117141_HeartModel-Lange-et-al_WhitePaper_LR.pdf

[2] Tsang W, Salgo IS, Medvedofsky D, Takeuchi MMD, Prater D, Weinert W, Yamat M, Mor-Avi V, Patel AR, Lang RM. Real-Time Automated Transthoracic Three-Dimensional Echocardiographic Left Heart Chamber Quantification using an Adaptive Analytics Algorithm. JACC Cardiovasc Imaging. DOI: https://doi.org/10.1016/j.jcmg.2015.12.020 CITATIONS: https://scholar.google.com/scholar?cites=4736404078278813501 "This new fully automated software has been recently validated and found to be reasonably accurate compared to manual 3D measurements using QLAB (3DQ) 2 in a group of over 150 patients at The University of Chicago. This promising software has the potential to enable the integration of 3DE volumetric LV and LA measurements into routine clinical workflows around the globe."

Philips HeartModel Cardiovascular 3D Ultrasound on EPIQ with Dr Ivan Salgo "5 Jul 2016"

https://www.youtube.com/watch?v=rJtnQk8A7aU

EPIQ 7G: Philipx Xmatrix has few button that might be related to AI but not quite sure.

AI AT Max PG
AI AT Vmax
AI DS Max PG
AI DS P1/2t
AI DS Vmax
AI End Dias PG
AI P1/2t Max PG
AI P1/2t Slope
AI P1/2t Time
AI P1/2t Vmax

Requires a bit of more investigation, perhaps the Manuals and data-sheets.

[mxochicale]

Samsung Medison’s NerveTrack™ automatic nerve tracking

• More accurate nerve detection: Not only does NerveTrack™ search for the nerve itself, but it recognizes the landmarks surrounding the nerve to improve detection accuracy. Using the Intel Distribution of OpenVINO toolkit’s CVAT (Computer Vision Annotation Tool), Samsung could increase the size of the image data set for training the NerveTrack™ feature by 7x, leading to improved accuracy of more than 20 percent.3

• Faster image processing and smoother workflows: One of the benefits of medical ultrasound imaging is the ability for users to interact dynamically with real-time data. The Intel® Core™ i3 processor and the Intel Distribution of OpenVINO toolkit can accelerate ultrasound image processing while simultaneously performing nerve detection. By identifying nerves in real time during ultrasonography, practitioners can minimize the possibility of complications while improving workflows https://www.intel.com/content/dam/www/public/us/en/documents/solution-briefs/samsung-nervetrack-solution-brief.pdf

• "Leveraging the Intel Distribution of OpenVINO toolkit for computer vision and annotation, Samsung Medison’s NerveTrack can potentially reduce scanning time by up to 30%."

• "NerveTrack was developed based on Intel’s OpenVINO toolkit. It uses inference to detect and identify the location of a nerve area in real time during an ultrasound scan, improving the treatment workflow for anaesthesiologists. To train Samsung’s real-time algorithm that automatically detects nerves in ultrasound images, a significant amount of clinical ultrasound data was required. And with Intel’s OpenVINO CVAT (Computer Vision Annotation Tool), the total volume of training data increased up to 7x, leading to improved accuracy of more than 20%. " https://www.med-technews.com/news/ai-and-vr-in-healthcare/samsung-medison-and-intel-tie-up-produces-ai-nerve-tracking-/