ROI-Aware AI Framework Improves Fetal Ultrasound Reconstruction
A new AI framework for fetal ultrasound reconstruction focuses on clinically relevant anatomical regions rather than global image quality. Researchers propose a two-stage ROI-aware representation learning method, instantiated for first-trimester nuchal translucency (NT) screening. A convolutional autoencoder first learns a globally faithful 128-D latent code using MS-SSIM, then refines the NT region of interest with intensity and edge constraints. Loss weights are calibrated automatically via gradient magnitudes. Tested under strict hospital-wise evaluation with one hospital held out, the method improves PSNR by up to +0.29 dB and reduces ROI MAE by up to 8.87%.
Key facts
- Proposed a two-stage ROI-aware refinement framework for fetal ultrasound reconstruction.
- Method targets first-trimester nuchal translucency (NT) screening.
- Uses a convolutional autoencoder (CAE) with MS-SSIM for global latent code.
- Refines NT ROI using L1 and normalized Sobel-edge constraints.
- Loss weights initialized via gradient-based calibration.
- Evaluated under multi-hospital domain shift with one hospital held out.
- Improves PSNR by +0.27 dB (val) and +0.29 dB (held-out test).
- Reduces ROI MAE by 8.87% (val) and 6.43% (held-out test).
Entities
—