Loading…
Conditional denoising diffusion model that synthesizes chest radiographs from patient demographics and echocardiographic ventricular measurements.
No providers recorded yet. Browse all providers
A chest radiograph costs a fraction of an echocardiogram, which is why models that screen for severe left ventricular hypertrophy (SLVH) and dilated left ventricle (DLV) from a plain film are attractive. The catch is that the only cohorts pairing the two modalities are shaped by who gets referred for an echo. In CheXchoNet the mean age is 62.3 years, SLVH prevalence is 8.6% and DLV prevalence 6.0%, and positives are scarcest in the under-60 group — precisely the population where catching structural heart disease early changes the most.
CXR-DDPM answers that by modelling the image distribution instead of a label. It is a conditional denoising diffusion probabilistic model whose conditioning signal is the echocardiogram itself: interventricular septal thickness (IVSd), posterior wall thickness (LVPWd) and internal diameter (LVIDd) at end-diastole enter as z-scored continuous values alongside age, sex and radiograph view. Because those measurements are continuous rather than binned into diagnostic classes, sampling is not confined to the combinations the training cohort happened to contain — an operator picks an age, a sex and a triple of wall and chamber dimensions, and the reverse diffusion process returns a radiograph consistent with them. Where cataloged chest-radiograph models such as CXR Foundation learn representations for downstream classification, this one generates the images those classifiers train on.
The model was built by Christopher Streiffer and colleagues at the University of Pennsylvania, across the Perelman School of Medicine, the Department of Radiology and the Center for Cardiovascular Informatics, and released as a medRxiv preprint in November 2024. The generator is the contribution; detection in the paper is performed by a separate DenseNet-121 CNN reproduced from prior work, used only to measure whether the synthetic images carry the signal.
The denoiser is a U-Net that takes a 224×224 single-channel grayscale image and predicts the noise added to it, built from downsampling and upsampling blocks with residual connections and attention layers, plus a timestep embedding and a state embedding carrying the conditional features. Noise is added on a cosine-beta schedule over 1,000 timesteps, and generation runs the full 1,000 reverse steps per batch. Training used the 90% patient-level partition of CheXchoNet — 64,277 posteroanterior radiographs from 22,220 patients collected at Columbia University Irving Medical Center between 2013 and 2018 — for 20 epochs at batch size 16 with AdamW at a learning rate of 1e-4 on a single NVIDIA A100. Diagnostic models trained on synthetic images alone reached a composite AUROC of 0.75 against 0.78 for real images; real data combined with positive synthetic samples reached 0.80 composite AUROC and 0.48 AUPRC, with the under-60 subgroup at 0.86. An independent 315-radiograph cohort from the University of Pennsylvania Health System checked the diagnostic models outside CheXchoNet.
The immediate use is enriching a screening dataset where it is thin: generating positive cases at a chosen age and sex tunes a diagnostic model toward a population the real cohort under-represents, which is where the reported gains concentrate. The same conditional sampling renders a radiograph from a patient's clinical feature vector, and offers a route to sharing image sets across institutions whose underlying records cannot travel.
The work extends the CheXchoNet line from detection into generation, and its specific contribution is conditioning on continuous clinical measurements rather than categorical labels, which is what makes prevalence and demographics tunable at sampling time. The authors state the limits plainly: no radiologist reviewed the generated images for realism, diffusion models are known to memorize training examples so the privacy argument is unsettled, and 1,000 forward passes per batch makes sampling expensive. On the Penn validation cohort — older on average than the target population — synthetic augmentation gave no improvement over real data alone. No trained checkpoint has been released; the repository provides training code, configuration files and a notebook, and CheXchoNet sits behind a PhysioNet data use agreement, so reproducing the model means retraining it.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.