Ultrasound Lung Aeration Map via Physics-Aware Neural Operators

TL;DR

Ultrasound lung imaging is limited by reliance on indirect B-mode artifacts and device-parameter variability. We introduce LUNA, a physics-aware Fourier neural operator that reconstructs 2D lung aeration maps directly from RF data, bypassing beamforming and B-mode interpretation. Trained on abundant simulated data generated with Fullwave-2 and fine-tuned on a small ex vivo dataset, LUNA achieves a mean aeration error of 9.4% on ex vivo swine lungs with a runtime of ~0.11 s per case, enabling ~9 Hz display. This yields a reader-independent, quantitative lung status map with potential for improved reproducibility and diagnostic utility across ultrasound devices and operator settings.

Abstract

Lung ultrasound is a growing modality in clinics for diagnosing and monitoring acute and chronic lung diseases due to its low cost and accessibility. Lung ultrasound works by emitting diagnostic pulses, receiving pressure waves and converting them into radio frequency (RF) data, which are then processed into B-mode images with beamformers for radiologists to interpret. However, unlike conventional ultrasound for soft tissue anatomical imaging, lung ultrasound interpretation is complicated by complex reverberations from the pleural interface caused by the inability of ultrasound to penetrate air. The indirect B-mode images make interpretation highly dependent on reader expertise, requiring years of training, which limits its widespread use despite its potential for high accuracy in skilled hands. To address these challenges and democratize ultrasound lung imaging as a reliable diagnostic tool, we propose LUNA, an AI model that directly reconstructs lung aeration maps from RF data, bypassing the need for traditional beamformers and indirect interpretation of B-mode images. LUNA uses a Fourier neural operator, which processes RF data efficiently in Fourier space, enabling accurate reconstruction of lung aeration maps. LUNA offers a quantitative, reader-independent alternative to traditional semi-quantitative lung ultrasound scoring methods. The development of LUNA involves synthetic and real data: We simulate synthetic data with an experimentally validated approach and scan ex vivo swine lungs as real data. Trained on abundant simulated data and fine-tuned with a small amount of real-world data, LUNA achieves robust performance, demonstrated by an aeration estimation error of 9% in ex-vivo lung scans. We demonstrate the potential of reconstructing lung aeration maps from RF data, providing a foundation for improving lung ultrasound reproducibility and diagnostic utility.

Figures (10)

• Figure 1: Overview: Luna (the Lung Ultrasound Neural operator for Aeration) reconstructs lung aeration maps from ultrasound radio-frequency (RF) data. a, The lung ultrasound (LUS) imaging process: Ultrasound devices scan the lung and the RF data is fed to Luna for reconstructing a lung air-tissue map, or aeration map, which is human-interpretable. The aeration map is used to estimate lung percent aeration, a critical clinical outcome for evaluation, monitoring and diagnostics lee2020lungkalkanis2022lung. Existing v.s. our approach: The current practice, B-mode lung image (upper right) requires manual interpretation of artifacts like B lines to assign LUS score volpicelli2012international, which has high variations due to the difficulty of identifying artifacts created by the complex wave propagation physics in the lung. LUS score is also coarse, image-level and semi-quantitative. Our approach (lower right): We reconstruct lung aeration maps that directly depict tissue-air maps, from which clinicians can directly read pixel-level aeration. The method is reproducible and provides quantitative two-dimensional aeration distribution. b,Luna performance on synthetic lung ultrasound data. Pixel-level aeration can be read from the predicted aeration map, which is visually similar to the true aeration map. c,Luna performance on real ex vivo lung ultrasound data. The average aeration prediction error is $9.4\%$, which well outperforms the sensitivity of the current scoring system.
• Figure 2: Luna architecture and training/fine-tuning data.a, Pipeline: During training, ultrasound RF data undergoes data augmentation and is processed by the physics-aware neural operator, Luna . Luna comprises two main components: a temporal Fourier neural operator, which processes the temporal dimension of RF data (corresponding to the output's depth dimension) and a spatial network, which captures lateral interactions in the RF data (corresponding to the output's lateral dimension). The model outputs a reconstructed aeration map, where each pixel value represents the aeration percentage. b, Generation of synthetic data, which is used for training Luna . Left: Combined aeration map comprised of the tissue-specifically segmented body wall (top) and underlying lung deformed to conform to its internal surface which models a realistic pleural interface. Middle: Stack of numerically simulated raw RF data of 128 transmit-receive events visualized as the amplitude of received backscattered signal in form receiver-time. Fullwave-2 pinton2021fullwave, a fullwave model of the nonlinear wave equation, is used as the simulation tool. 10k samples are used to train Luna . c, Acquisition of real data, which is used for training Luna . Scanning of fresh porcine lungs of known aeration (displacement method) through chest wall fragment in the water tank (ex vivo) using a programmable ultrasound machine and linear transducer. 18 samples are used to fine-tune Luna .
• Figure 3: Visualization of lung ultrasound B-mode images and aeration maps, in silico and ex vivo.a, Ground-truth and reconstructed aeration maps in in silico, showing close visual similarity. Aeration maps are flattened for visualization purposes. b/c, Reconstructed aeration maps overlaid on B-mode images for in silico (b) and ex vivo (c) experiments. The reconstructed aeration maps closely align with the B-mode images, effectively capturing aeration changes. We also provide in silico ground-truth aeration maps to which the predictions are similar (b). The lower right plot in each subfigure depicts the reconstructed 1D percent aeration curve.
• Figure 4: Luna performance, in silico and ex vivo.a,Luna percent aeration prediction error of in silico and ex vivo data. The error gap between them is 4.2%. b,Luna percent aeration prediction error grouped by ground-truth percent aeration and chest wall depth, in silico and ex vivo. Both sets have consistent errors for different aeration and chest wall depth, indicating the Luna 's robustness to such lung characteristics. c,Luna aeration map 2D reconstruction performance (PSNR and SSIM) grouped by ground-truth percent aeration (left two subfigures) and chest wall depth (right two subfigures) in silico. The performance is consistent for samples with different percent aeration and chest wall depth.
• Figure 5: Architecture design and training details of Luna .a, Chest wall segmentation network: This network reconstructs the chest wall segmentation from the RF data. b, Fourier feature learning allows invariance learning to the temporal delay (with details in Section \ref{['sec:architect']}). c, Temporal augmentation: The temporal dimension ($t=0$ to $t=200$) is randomly masked during training, excluding the initial signal corresponding to chest wall structures, which are irrelevant to aeration reconstruction. d, Frequency-space processing: Luna operates directly in the frequency domain, where Fourier features provide invariance to temporal delays in the RF data and variations in chest wall depth. e, Calibration plot: Luna demonstrates strong calibration, with predicted aeration percentages closely matching true values, ensuring accurate and reliable predictions.