01

1. Capture an existing ultrasound examination

The 2024 study used five-second parasternal long-axis videos recorded from the display through a smartphone application. This separates ultrasound image formation from submission to AI: the ultrasound machine creates the image, and the phone captures and transmits the clip.

Display capture introduces variations such as reflections, hand movement, and framing. These affect the input, which is why quality screening precedes classification instead of assuming that every recorded clip is suitable.

Source: 2024 paper, pp. 127–128; CNF 2026, pp. 14, 26–34

02

2. Assess quality before cardiac classification

The paper describes checks for obscuring reflections, missing or multiple ultrasound regions, an image region that is too small, insufficient duration, and excessive motion. This stage separates acquisition problems from the cardiac-function classification itself.

Quality also defines the study population. The paper reports 1,336 submitted videos and 923 remaining after screening. Its accuracy therefore describes the included data and should not be extended to excluded clips.

Source: 2024 paper, pp. 128–130; CNF 2026, p. 45

03

3. Prepare the video for the model

The published pipeline uses an ultrasound-region detector adapted from DAMO-YOLO, ORB features to help stabilize the video, and CLAHE with thresholding to adjust lighting. These steps make the visual input more consistent before it reaches the classifier.

These details describe the system studied in 2024 and do not establish that every later version uses identical code or models. Evaluation of a new version needs to identify its version, preprocessing pipeline, and test data.

Source: 2024 paper, p. 129

04

4. Classify video and consolidate outputs

The paper describes an adapted ResNet-101 with video classification, initially using five categories and then postprocessing the output into Reduced EF, Mildly Reduced EF, and Preserved LV. This is category classification, not evidence of precise continuous EF measurement at every value.

Performance depends on the number of categories: the paper reports 68.5% for five categories and 96.2% for three on the test set. Combining categories means some errors between subcategories no longer count as three-category errors. Every percentage should therefore identify the classification task.

Source: 2024 paper, pp. 129–131; CNF 2026, p. 44

05

5. Return the result for clinical review

The architecture slide connects the application over the internet to an Echo AI Server and API. Report examples illustrate printing and transfer. The project plan develops these into a Mobile Application, Cloud-based AI Platform, Backend API, and Dashboard / Clinical Report.

CNF page 45 reports processing in under three seconds in the presentation context, but does not provide a complete end-to-end timing protocol. It should not be treated as a speed guarantee across all networks and devices.

Source: CNF 2026, pp. 33–35, 43, 45; EasyEF project brief, 22 June 2026

06

6. Device integration and additional models

The sources propose handheld ultrasound integration and an Echo AI Box concept for existing machines. CNF page 60 also seeks collaboration on additional abnormalities, including valve and pericardial conditions. These are extension directions, not a list of fully validated capabilities.

Some conditions proposed for future work were excluded from the 2024 study. Applying its 96.2% result to detection of those conditions would therefore exceed the evidence.

Source: CNF 2026, pp. 46–47, 60; GPO 2026, pp. 20–23

SOURCE NOTES

Sources for this page

  1. บทความวิจัยปี 2567 / 2024 original article — J Prapokklao Hosp Clin Med Educat Center 41(2):123–132 ↗
  2. เอกสารนำเสนอ / Presentation: EasyEF Lunch Symposium, CNF 17th, 26 March 2026 — หน้า / pp. 14, 26–35, 43–47, 60
  3. เอกสารนำเสนอ / Presentation: EasyEF AI Echo, GPO, 4 February 2026 — หน้า / pp. 20–23
  4. ข้อเสนอโครงการ EasyEF ฉบับ 22 มิถุนายน 2569: ขอบเขตและแผนงาน / EasyEF project brief, 22 June 2026: scope and planned work

Presentations reflect information at the time they were prepared. Explore context and evidence status in the resource library.

Explore the source library →