A Clinician-Led Feasibility Study for Echocardiographic View Classification Using Standard Workstation Hardware and Routine Clinical Archives.

PURPOSE: To evaluate whether a frontline echocardiographer, without engineering support or dedicated research hardware, can independently develop a validated five-class echocardiographic view classifier using artificial intelligence-assisted coding, standard departmental workstation hardware, and routine clinical archives.

METHODS

A clinician used artificial intelligence-assisted coding (Kimi 2.6, Moonshot AI) to develop a pipeline on a standard Tomtec echocardiographic analysis workstation (Intel Core i7, NVIDIA RTX 4060 8 GB VRAM, 32 GB RAM) running Ubuntu 22.04 LTS via Windows Subsystem for Linux 2. From 1387 transthoracic echocardiograms, 9796 B-mode end-diastolic frames were extracted. A batch annotation strategy was employed: 466 seed images were annotated to train an overfitted EfficientNet-B0 screener, which pre-classified 5029 high-confidence standard-view frames. These were batch-reviewed in approximately 24 min, yielding 5028 verified frames from 1242 patients, partitioned into training (n = 3545), validation (n = 731), and internal test (n = 752) sets. A five-class ResNet-18 classifier was trained and evaluated on the primary dataset and evaluated on an external validation dataset (74 patients, 263 frames) without retraining.

RESULTS

The total expert annotation burden was 32 min. The classifier achieved 98.27% accuracy (95% CI: 97.06%-99.08%) on the internal test set with a macro F1 of 0.9784. On external validation, overall accuracy was 71.48% (95% CI: 65.61%-76.86%) with a macro F1 of 0.7100. Parasternal short-axis and apical two-chamber views generalized robustly (100% and 94.23%, respectively), whereas apical four-chamber, apical three-chamber, and parasternal long-axis views showed lower accuracy due to device-specific image processing differences.

CONCLUSION

Clinical artificial intelligence development in echocardiography is substantially more accessible than conventionally assumed. A single clinician using existing hardware and artificial intelligence-assisted coding can produce a validated view classifier with minimal annotation burden, providing a practical roadmap for in-house customization.