Automated view selection tool delivers high accuracy across standard and handheld echocardiography: results from a government-funded project
D D Pelin, G Schummers, A Calvarons, S Zoernack, N Hitschrich, M LeMaire, L Krieger, M Schreckenberg, K Franitza, S Franz, B Lengenfeld, K Hu, S Stoerk, C MorbachAbstract
Background
Advances in ultrasound technology have produced compact, affordable handheld devices for point-of-care imaging, albeit with some loss of resolution. In parallel, machine-learning tools promise to reduce reader variability and improve reproducibility. A major bottleneck remains view selection, which marks the first step in the reading process.
Aims
We evaluated the performance of an automated view selection tool (AVST) applied to apical 4-, 2-, and 3-chamber (CH) views acquired on a console-based ultrasound system (CBUS) and on a handheld ultrasound device (HUD). Accurate identification of these views is essential for the derivation of automated standard measurements like global longitudinal strain (GLS).
Methods
This government-funded project is a collaboration between a tertiary care university hospital and a medical imaging softwarecompany that uses a federated machine-learning framework. We recruited 2043 patients (mean age 64±16 years, 44% women) referred for transthoracic echocardiography. After a high-resolution study on a CBUS , trained sonographers acquired images with a HUD according to a pre-specified protocol. For the current analysis, we selected 300 consecutive patients (63.8±15.5 years, 43% women) from the project cohort. The AVST automatically labelled candidate apical views, which an expert reader either accepted or declined; if no suitable view was found, the reader searched the original study. All labeled images were loaded into the analysis platform of the local Academic CoreLab for further measurements.
Results
Among 2,783 HUD B-mode clips and 3,852 CBUS B-mode clips, the AVST identified 1,404 and 1,672 clips, respectively, as appropriate apical views. The expert reader declined 7 HHD studies (2.3%) and 12 CBUS studies (4 %) (Table). Reasons for CBUS declines encompassed software bug (n=4), failure to detect 3CH view (n=2) or 2CH view (n=1), non-acquisition of 3CH view (n=2), motion artifacts (n=3); HUD declines were due to insufficient image quality (n=6), or motion artifact (n=1). The remaining 293 HHD and 288 CBUS studies were accepted by the expert reader; these clips were correctly classified and automatically analysed for GLS with no further manual intervention.
Conclusion
AVST performed equally well on console and handheld echocardiography. This multi-vendor view selection approach appears feasible and reliable, enabling user-independent GLS measurement. Its implementation in routine practice will require prospective diagnostic studies.