Registration of interest

Research access to AURORA, and a consortium to train AURORA 2.0

AURORA is a convolutional neural network that scores routine 2D transabdominal ultrasound for placenta accreta spectrum (PAS). It was developed at Baylor College of Medicine and Texas Children's Hospital. We invite clinicians and scientists who want to evaluate it, and centers with pathology-confirmed PAS cases that want to help train the next version.

How AURORA scores a scan

DICOM cine clip2D transabdominal views of the lower uterine segment and placenta.
Frame preprocessingGrayscale, crop to the image region, resize to 256 × 256.
CNN per frameEfficientNetB0 (ImageNet-pretrained) outputs a probability for every frame.
Patient scoreFrame probabilities are averaged across all of a patient's clips.
Clinical ensemblePatient score plus prior cesarean count and previa, combined by logistic regression, random forest and gradient boosting.

Figure 1. AURORA 1.0 pipeline. Grad-CAM maps show which image regions drove each frame's score.

Performance on the held-out test set

Seventeen test patients (9 PAS, 8 without PAS) from the development center. The intervals are wide, and external validation at other centers is the next step.

Ensemble (CNN + prior cesareans + previa)CNN alone

Sensitivity

100.0 (66.4–100)
88.9 (51.8–99.7)

Specificity

75.0 (34.9–96.8)
75.0 (34.9–96.8)

PPV

81.8 (48.2–97.7)
80.0 (44.4–97.5)

NPV

100.0 (54.1–100)
85.7 (42.1–99.6)

Accuracy

88.2 (63.6–98.5)
82.4 (56.6–96.2)

AUC

0.972 (0.875–1.000)
0.861 (0.633–1.000)

Figure 2. Point estimates with 95% confidence intervals on a 0 to 100% scale (AUC on 0 to 1). Proportions use exact Clopper-Pearson intervals. Ensemble confusion matrix: 9 true positives, 6 true negatives, 2 false positives, 0 false negatives.

Ways to participate

Research use

Evaluate AURORA at your institution

A browser-based research viewer is in preparation; scans are processed on your own computer and never uploaded.
Access is granted to registered researchers under a research-only license.
Output is a continuous research score, not a diagnosis.

Data contribution

Help train AURORA 2.0

For centers with histopathology-confirmed PAS cases.
Images are de-identified at your site before anything is shared.
Participation runs under institutional IRB approval and signed data agreements.
Contributors are credited under a consortium authorship policy.

Register

The form asks for your name, email, institution, role and country; whether you want to use AURORA, contribute data, or join a multicenter collaboration; your ultrasound systems and reporting software; your approximate annual number of PAS cases; and whether PAS is confirmed by histopathology at your center.

Open the registration form

QR code linking to this registration page

Scan to open this page on a phone.