EchoGuide AI: An Uncertainty-Aware Deep Learning Pipeline for Echocardiographic View Classification and Cardiac Structure Segmentation, with an Evaluation of Label-Free Quality Scoring and of the Data Barrier to Probe Guidance

Background. Transthoracic echocardiography depends heavily on the operator. A recent multi-task framework proposed deriving image quality from the predictive uncertainty of view-classification and segmentation networks, and recommending corrective probe maneuvers [1]. Its authors reported training and validating the system on a 43-case maneuver-tagged dataset that they constructed; we did not find a public release of these data, which prevents independent reproduction of the maneuver component. Objective. To build an open, reproducible counterpart on public data and to establish which parts of such a system public data can actually support. Methods. We used the public CAMUS dataset [2]: 500 patients and 1,998 expert-annotated apical two- and four-chamber (A2C, A4C) frames at end-diastole and end-systole. We partitioned it at subject level into 300/100/100 patients with a fixed seed and an automated leakage check. View classifier: an ImageNet-initialised ResNet-18 separating A2C from A4C. Segmentation network: a U-Net trained from scratch to segment the left-ventricular (LV) cavity, myocardium (MYO) and left atrium (LA). Quality score: a score Q built from normalised softmax entropy, with no quality labels used in training. We compared Q with independent expert quality ratings. Statistics: all test-set statistics use 95% confidence intervals (CIs) from a subject-level cluster bootstrap. Results. On 400 frames from 100 held-out patients: View classification: accuracy 0.985 (95% CI 0.9675–0.9975), macro-F1 0.985, ROC-AUC 0.996 (0.989–1.000). The ex-pected calibration error (ECE) was 0.016. Five of the six errors were A2C frames predicted as A4C, and five of the six erroneous frames had been rated Good by the expert. Segmentation (Dice): LV 0.934 (0.926–0.940), MYO 0.873 (0.865–0.879), LA 0.892 (0.880–0.904), macro 0.900 (0.893–0.906). Segmentation (95th-percentile Hausdorff distance, HD95): 4.01, 4.23 and 5.42 mm for LV, MYO and LA. Quality score: Q correlated moderately with the model’s own per-frame Dice (Spearman ρ = 0.554) but only weakly with expert ratings (ρ = 0.281). ROC-AUC for detecting Poor frames was 0.683, and 94.8% of frames scored ≥ 90 even though experts rated 54.5% as Good. Probe maneuvers: none of the four public datasets we reviewed (CAMUS, EchoNet-Dynamic, TMED-2, HMC-QU) provided the probe-maneuver annotations required for supervised maneuver prediction; CAMUS contains none (0 of 1,998 frames). The maneuver component was therefore not trained, and no maneuver result is reported. Conclusions. On public data, view classification and multi-structure segmentation can be reproduced with a transparent, leakage-controlled protocol. Single-pass entropy, however, tracks model error far better than expert-judged image quality. For probe guidance, the absence of maneuver annotations in the reviewed public datasets, rather than model design, was the limiting factor. The models are served by a full-stack research application that exposes its supported scope explicitly and never fills unsupported outputs.