Convolutional networks
Can a neural network recognise which heart-ultrasound view it's seeing?
Open access · cc by · source: Europe PMC
A convolutional neural network named 15 standard echocardiogram views more accurately than board-certified experts, using tiny low-resolution images.
Study at a glance
- Design
- Computational / modelling — Supervised CNN trained on 15-view labelled echocardiogram frames, with train/validation/test split by patient study, a random-label negative control, and a comparison with 4 human experts.
- N
- N=267 · Echocardiogram studies from different patients: 240 for training and validation (213 train, 27 validation) plus 27 for testing, yielding 223,787 labelled frames; 4 board-certified echocardiographers each classified 1500 test images.
- Population
- Randomly selected clinical transthoracic echocardiograms from UCSF, 2000-2017, adults aged 20-96
- Outcome
- Per-view and overall classification accuracy, F-score and AUC on held-out images and videos; comparison with expert accuracy
Structured fields used in claim comparison tables when every cited study has a complete layer.
Key findings
On videos (classified by majority vote across frames) overall test accuracy was 97.8%. On single still images it was 91.7%, compared with 79.4% for the human experts on a subset of the same images. Training on randomly shuffled labels gave 6.9% accuracy, roughly chance for 15 classes. Errors mostly involved genuinely similar views such as apical two- versus three-chamber, and masking the heart itself, but not other image regions, sharply reduced accuracy.
Methodology
The authors took 240 randomly chosen clinical echocardiogram studies for training and validation and 27 more for testing, had an expert label every clip by view, and shrank each frame to a 60 × 80-pixel greyscale image. They trained a VGG-inspired convolutional network on 15 views, keeping training, validation and test images from different patients. They checked the model against random-label training, compared it with 4 board-certified echocardiographers on the same test images, and used occlusion and saliency maps to see which parts of the image it relied on.
Limitations
All data came from one medical centre, labelled by a single echocardiographer, so there is no external validation or measure of label noise. The human comparison used only 4 experts viewing the same shrunken low-resolution stills the model saw, a format they do not normally work from, which may understate human skill. View classification is only a first step; the paper does not show the model improves diagnosis or patient outcomes.
How this study connects
Role on claims
Each row is a claim on a concept or method page where this paper supports, challenges, or qualifies the statement. Roles are hand-checked — not a model guess.
CNNs can learn genuine image structure for medical view recognition.
On clinical echocardiograms from one US centre, a CNN classified 15 views with 97.8% accuracy on videos and 91.7% on single stills, versus 79.4% for four experts on the same low-resolution stills; with shuffled labels it fell to 6.9%.
Evidence for the claim as stated.
CNNs can learn genuine image structure for medical view recognition.
On clinical echocardiograms from one US centre, a CNN classified 15 views with 97.8% accuracy on videos and 91.7% on single stills, versus 79.4% for four experts on the same low-resolution stills; with shuffled labels it fell to 6.9%.
Scope note — Single centre, single labeller, no external validation; experts saw an unfamiliar image format.
Limits the claim's scope: a different population, assay, or outcome.
Image tasks with clear visual structure (echocardiogram views, tumour histology) produced very high accuracies, while molecular and protein tasks gave moderate or split-dependent results; these are different data types and evaluation setups, not a direct conflict.
Evidence for the claim as stated.
Open questions
Tensions this paper is part of
From concept pages' “where studies disagree.” Disagreement means the same question; scope means different assays, populations, or outcomes.
Image tasks with clear visual structure (echocardiogram views, tumour histology) produced very high accuracies, while molecular and protein tasks gave moderate or split-dependent results; these are different data types and evaluation setups, not a direct conflict.
Related papers in this topic
Same topic cluster — not a recommendation engine.