Abstract
Synchrotron X-rays can be used to obtain highly detailed images of parts of the lung. However, micro-motion artifacts induced by such as cardiac motion impede quantitative visualization of the alveoli in the lungs. This paper proposes a method that applies a neural network for synchrotron X-ray Computed Tomography (CT) data to reconstruct the high-quality 3D structure of alveoli in intact mouse lungs at expiration, without needing ground-truth data. Our approach reconstructs the spatial sequence of CT images by using a deep-image prior with interpolated input latent variables, and in this way significantly enhances the images of alveolar structure compared with the prior art. The approach successfully visualizes 3D alveolar units of intact mouse lungs at expiration and enables us to measure the diameter of the alveoli. We believe that our approach helps to accurately visualize other living organs hampered by micro-motion.
Introduction
Visualizing pulmonary alveoli—the microscopic air sacs responsible for gas exchange—within a living organism presents a significant challenge. Operating at the micrometer scale, these structures are highly susceptible to "motion artifacts" caused by the heartbeat and respiration, which blur images and hinder quantitative analysis. Historically, researchers have relied on invasive imaging, which alters internal lung pressure, or chemically "fixed" tissue samples, which suffer from significant shrinkage and distortion.
While synchrotron X-ray Computed Tomography (CT) offers the necessary brightness for high-resolution imaging, reconstructing clean 3D volumes from noisy, motion-blurred projections remains difficult, particularly when ground-truth data is unavailable. This paper introduces an unsupervised deep-learning framework that leverages a "Deep Image Prior" (DIP) approach with interpolated latent variables to reconstruct high-resolution 3D images of lung alveoli in living, intact mice.
Methodology
Experimental Setup and Expiratory Gating
To minimize motion during data acquisition, the researchers synchronized a mechanical ventilator with the synchrotron X-ray system using transistor-transistor logic (TTL) signals. This "expiratory gating" ensures that X-ray projections are only captured at the end of the expiration phase, a period when lung motion is naturally at its minimum.
Deep Learning Reconstruction Framework
The core of the reconstruction method is an unsupervised neural network that does not require labeled training data. It extends the standard Deep Image Prior (DIP) by incorporating interpolated latent variables to handle 3D volume reconstruction efficiently.
The approach uses the inductive bias of convolutional neural networks to represent structured images while filtering out noise. To reconstruct the 3D volume using a 2D network (optimizing GPU memory), the vertical z-axis is encoded as a sequence of latent variables. Piece-wise linear interpolation between these variables forces the network to maintain structural continuity between adjacent slices. The network is trained by minimizing the difference between its generated images (after passing through a Radon transform) and the raw X-ray projections, supplemented by a Total Variation (TV) regularizer.
Datasets
The study utilized two primary datasets for validation:
- Real Synchrotron Dataset: X-ray projections from five live, intact eight-week-old nude mice, captured at the SPring-8 facility in Japan.
- Synthetic Dataset: Created from manually labeled binary images of alveolar structures with added Gaussian and speckle noise to quantitatively evaluate structural recovery.
Results
The proposed method demonstrated significant improvements over traditional algorithms like Filtered Back Projection (FBP) and Conjugate Gradient for Least Squares (CGLS).
- Quantitative Performance: On the real dataset, the method achieved a pixel accuracy of 91.15% and a higher Jaccard index than baselines. In synthetic tests, it maintained a Peak Signal-to-Noise Ratio (PSNR) of ~32.9 dB, compared to ~11.6 dB for FBP.
- Biological Findings: The high-quality 3D reconstructions revealed that the diameter of alveoli in intact, living lungs is approximately 79 µm—nearly twice the size typically reported in studies using fixed or deflated tissue (~40 µm). This suggests that previous understandings of lung mechanics may be based on distorted data.
Discussion and Limitations
While successful, the framework has limitations. It assumes stable, periodic motion; irregular movements (e.g., due to light anesthesia) can cause reconstruction failure. Additionally, the "Center of Rotation" (CoR) for the CT scan currently requires manual adjustment, though future work aims to automate this as a learnable parameter. Finally, the computational cost is notable, with a single 3D volume reconstruction taking approximately 30 minutes on a high-end GPU.
Conclusion
This research provides a robust tool for the non-invasive, high-resolution visualization of lung microstructures in vivo. By combining hardware synchronization with a novel unsupervised deep-learning architecture, the method overcomes the hurdles of micro-motion artifacts, offering new insights into the true morphology of living organs.
Source: Scientific Reports (2023) 13:1738 | https://doi.org/10.1038/s41598-023-27627-y
Summary updated