CPSNet uses physics-inspired AI to detect lung ultrasound B-lines without labels

A team of researchers (Tianqi Yang, Oktay Karakuş, Nantheera Anantrasirichai, Marco Allinovi, Alin Achim) has introduced CPSNet, a physics-inspired deep unfolding framework that detects B-lines in lung ultrasound images without requiring labeled training data. B-lines are vertical artifacts indicating pulmonary edema or other lung conditions, and their automated detection is critical for rapid triage. CPSNet works by unfolding a Cauchy proximal splitting algorithm into a forward-backward style neural network, using skip connections to iteratively clean noisy Radon domain images. The team also developed a novel Radon-Based Neighbor Reconstruction Loss that enforces consistency between subsampled and reconstructed image pairs, with regularization to handle noise. Because CPSNet trains in an unsupervised manner (no ground truth images), it avoids the costly and time-consuming manual annotation that plagues supervised methods.

Evaluated on lung ultrasound data, CPSNet demonstrated strong performance across structural similarity index (SSIM), proxy peak signal-to-noise ratio (PSNR), and relative L2 norm. In B-line detection tasks, it showed greater stability, adaptability, and efficiency compared to both traditional signal-processing approaches and modern object detection models. Crucially, it preserved line structures while minimizing false positives—a common problem in clinical settings. The authors conclude that CPSNet offers a reliable, efficient solution for lung ultrasound diagnostics, supporting faster and more accurate clinical decision-making. The code is publicly available on GitHub, and the paper appears in Digital Signal Processing (2026).