Proof of Concept
Real-Time Fetal Heart Detection
A two-step AI pipeline that draws a color-coded guidance overlay on a live ultrasound feed to help sonographers capture high-quality diagnostic images of the fetal heart.
On 28 July 2026 we tested one of the results on this page and it did not hold up. The accuracy figures for the image quality classifier, Step 2 of the pipeline, have been withdrawn. They measured the model recognising the artificial degradations we applied when generating its training labels, not its ability to judge whether an ultrasound image is diagnostic.
The heart detection results, Step 1, are unaffected and were measured separately. The training set size described on this page was also corrected in July 2026.
Full method and evidence: Experiment 17, Validating the Image Quality Classifier. For how our process changed as a result, see our retrospective re-assessment.
Background: Congenital Heart Defects (CHD) are among the most common birth defects, affecting nearly 1 in 100 births. Effective prenatal detection depends on acquiring high-quality ultrasound images of four standard cardiac diagnostic planes. Sonographer fatigue and variability in image quality remain significant barriers to consistent early detection.
Objective: To build and validate a proof-of-concept real-time guidance tool that detects the fetal heart in a live ultrasound feed, assesses the quality of the captured view, and displays a color-coded bounding-box overlay (GREEN, YELLOW, or RED) to guide the sonographer toward capturing a diagnostically useful image. The POC targets the four-chamber view (4CV) using the publicly available FOCUS dataset.
Approach: A lightweight two-model pipeline was designed to run entirely on consumer CPU hardware (no GPU required) to minimise deployment cost and maximise accessibility for under-resourced clinical settings.
The inference pipeline is deliberately minimal: three sequential steps that together complete in approximately 80-150 ms per frame on an Intel i7-8665U (no GPU).
- YOLO11n: Heart Localisation. A nano-scale YOLO model detects the fetal heart region and outputs a bounding box. Trained using 5-fold cross-validation on Kaggle T4 GPUs, on 267 images from the FOCUS dataset. (An earlier version of this page described FOCUS and the Kaggle four-chamber dataset as two separate training sources. They are the same images. See the correction published July 28, 2026.)
- EfficientNetV2-S: Image Quality Assessment. The region cropped by YOLO is passed to a fine-tuned EfficientNetV2-S classifier that assigns one of three quality labels: GREEN (diagnostically suitable), YELLOW (marginal), or RED (inadequate). Trained on 1,160 crops derived from 232 source images in the FOCUS dataset. (This page previously said 1,500 crops, and did not mention how few original images they came from. Each source image yields one GREEN crop plus four artificially degraded ones, so the crop count overstates how much independent data stands behind this model.)
- OpenCV: Real-Time Overlay. The bounding box is rendered on the live frame in the colour corresponding to the IQA output, giving the sonographer immediate visual feedback.
The video below demonstrates the POC pipeline running against sample frames from the FOCUS test set, showing the three quality states firing in real time.
Results
| Metric | Value |
|---|---|
| mAP@50 (mean, 5-fold CV) | 0.582 ± 0.141 |
| mAP@50 (best fold: Fold 2) | 0.699 |
| Precision (mean) | 0.711 |
| Recall (mean) | 0.648 |
These figures were withdrawn on 28 July 2026.
Testing showed they reflect detection of our own synthetic image degradations rather than assessment of diagnostic quality. A logistic regression on nine low-level image statistics scored higher on the same task, and across 775 real ultrasound crops the model never returned a RED rating. See Experiment 17.
| Metric | Value |
|---|---|
| Val Accuracy (mean, 5-fold CV) | WITHDRAWN (was 99.89%) |
| Test Accuracy (overall) | WITHDRAWN (was 97.62%) |
| Test Accuracy (GREEN) | 88.10% |
| Test Accuracy (YELLOW) | 100.00% |
| Test Accuracy (RED) | 100.00% |
Key Findings
The YOLO11n localiser achieved a best-fold mAP@50 of 0.699 on limited training data. These are internal research results under controlled conditions, not measures of clinical performance.
The quality classifier's accuracy figures were withdrawn on 28 July 2026. This page previously reported 99.89% validation accuracy and 97.62% test accuracy, including perfect recall on the RED class. Testing showed those numbers reflect the model recognising the artificial blurring we applied when generating its training labels, rather than judging diagnostic quality. A basic statistical method using nine simple image measurements scored higher on the same task. Across 775 real ultrasound images the model returned RED zero times. Full detail: Validating the Image Quality Classifier.
Limitations
The current POC covers only the four-chamber view (4CV) and has been validated solely on the FOCUS dataset. The quality labels used to train and evaluate the IQA classifier were generated programmatically rather than rated by clinical experts, which likely inflates its accuracy figures; validating the quality signal against expert ratings is our first priority on the roadmap. YOLO localisation performance shows moderate variance across folds (± 0.141), suggesting the model would benefit from a larger and more diverse training set. CPU-only inference also limits frame rate in a live clinical setting.
Added July 28, 2026. An audit of our own data added three limitations to this list, and they matter more than the ones above.
First, every number on this page comes from 267 images taken from a single dataset and a narrow range of ultrasound machines. That is a much smaller evidence base than this page previously implied.
Second, the variance figure above may be a limit of our measurement rather than of our model. With roughly 53 images used for scoring in each round, we do not currently know the smallest difference our testing is capable of detecting. Until we do, our model comparisons should be read with caution, and we are running a dedicated experiment to establish it.
Third, the model has never been tested on an ultrasound machine it was not trained on. Performance on unfamiliar equipment is the single most relevant question for a tool meant to work in under-resourced settings, and we have not yet answered it. We expect the honest answer to be worse than the figures above, and we plan to publish it either way.
Next Steps
Phase 4 will add a per-view capture checklist to the inference UI, with 4CV as the active slot and placeholders for LVOT, RVOT, and 3VV. Phase 5 will expand coverage to all four ISUOG standard cardiac planes using the SonoNet dataset, and will include TensorRT / OpenVINO optimisation for real-time GPU-accelerated deployment and clinical validation with expert sonographers.