cGAN boosts low-end POCUS image quality by 86% in new benchmark dataset
First paired dataset of low-end vs high-end ultrasound enables 54% SSIM lift with deep learning.
A team led by Lennard M. van Karnenbeek introduced the first accurately paired point-of-care ultrasound (POCUS) dataset — POCUS-IQ — built with a custom automated gantry to align low-end handheld scans with high-end clinical ultrasound images. The dataset contains 1,064 paired ex vivo tissue and phantom images, enabling direct benchmarking of image enhancement algorithms.
The team applied a conditional GAN based on pix2pix with a U-Net generator, using both L1 and SSIM losses and pretraining on simulated data. Results: SSIM jumped from 0.29 to 0.54, PSNR from 19.16 dB to 22.41 dB, and no-reference metrics NIQE and PIQE improved by 44% and 36% respectively. The framework shows promise for turning low-cost handheld ultrasound into a diagnostically viable tool in low-resource settings.
- First publicly available paired dataset (POCUS-IQ) of low-end handheld vs high-end ultrasound images, 1,064 pairs.
- cGAN (pix2pix + U-Net) boosted SSIM from 0.29 to 0.54 (+86%) and PSNR from 19.16 to 22.41 dB.
- No-reference quality scores dropped sharply: NIQE 7.95→4.44, PIQE 31.12→19.99, indicating perceptual improvement.
Why It Matters
Enables low-cost POCUS devices to match high-end diagnostic quality, expanding ultrasound access in underserved areas.