LungViT

Functional small airways disease is among the earliest measurable signs of airflow obstruction in COPD, and the only established way to image it in vivo needs two breath-hold CT scans: one at total lung capacity (TLC) and one at residual volume (RV), co-registered and thresholded by parametric response mapping. The expiratory half is the bottleneck. It adds dose and scan time, it needs a technologist to coach the patient to a reproducible exhale, and no harmonized protocol exists for doing so, so it is rarely acquired outside research cohorts. Every inspiratory-only chest CT — including every lung cancer screening scan — therefore carries small airways information that cannot be read out.

LungViT synthesizes the missing scan. It is a conditional generative adversarial network from the University of Iowa, with collaborators at UCLA and the University of Utah, that translates inspiratory CT intensities into the corresponding expiratory intensities across the whole lung. Because the output is a full CT volume in the inspiratory scan's own coordinate frame, parametric response mapping runs on it unchanged: the model slots in ahead of an established measurement rather than replacing it with an opaque score.

The architecture targets two failure modes of earlier CT generators: slice-based 2D networks leave visible discontinuities across the axial and sagittal planes, and 3D models small enough to fit in GPU memory blur away the vessels, fissures and parenchymal texture that density thresholds depend on. It sits alongside volumetric chest CT generators such as Lung-DDPM, which synthesizes scans from a semantic layout rather than translating a paired one.

#Key Features

• Small airways disease from one acquisition: Thresholding the real inspiratory scan against the synthetic expiratory scan yields an fSAD percentage without an RV acquisition.
• SwinSEER generator: A four-stage shifted-window (Swin) transformer encoder feeds a convolutional decoder whose squeeze-and-excitation blocks recalibrate feature channels at every resolution — a SwinUNETR variant built for dense regression.
• Multiview texture loss: A 3D extension of the DISTS perceptual metric, computed across orthogonal views, supervises texture and style transfer, alongside a multiresolution voxel-consistency loss at the last three decoder layers.
• Ensemble cascading: Two generators are chained, the second conditioned on the first's output and sharing no weights with it, which measurably sharpens the refined volume.
• Volume-scale output: Trained on 128-voxel cubic patches and assembled at inference into volumes of 320 × 320 × 320, well beyond the reach of most 3D GANs.

#Technical Details

LungViT is trained as a volumetric conditional least-squares GAN on paired patches drawn from co-registered TLC/RV volumes. The training set is 1,055 SPIROMICS participants spanning never-smokers and GOLD stages 0 through 4; RV volumes are matched to TLC by a mass-preserving deformable registration constrained on fissure structure (Dice 0.988, average symmetric surface distance 0.467 mm), then resampled to 1 mm isotropic and clipped to [-1024, 1024] HU. Implementation is PyTorch and MONAI on a single NVIDIA A100 at batch size 4, with Adam and imbalanced generator/discriminator learning rates of 2 × 10⁻⁴ and 5 × 10⁻⁵. On a disjoint 512-subject SPIROMICS test set, the cascaded model reaches 24.75 dB PSNR, 0.842 SSIM, 0.096 LPIPS and 3.26% normalized mean squared error, ahead of Pix2Pix, SAGAN, PTNet, Vox2Vox and adversarially trained UNETR and SwinUNETR baselines on most metrics. Bland-Altman analysis puts the bias in derived fSAD at 2.33% and in air trapping at 1.56%, both underestimates. Generalization was checked on 200 COPDGene subjects scanned under a different protocol.

#Applications

A follow-up study applied the trained generator, without refitting, to 1,458 held-out SPIROMICS participants and 458 from COPDGene. The resulting single-volume measure, fSAD_TLC, correlated with conventional dual-volume fSAD at R = 0.895 and R = 0.897 respectively, and was independently associated with FEV1, FEV1/FVC, St. George's Respiratory Questionnaire score and five-year FEV1 decline in both cohorts after adjustment for emphysema and smoking history. It also proved more repeatable across 2–6 week rescans than the two-scan measure (intraclass correlation 0.99 against 0.83), since expiratory effort varies between visits in a way inspiratory effort does not. The payoff is retrospective: cohorts and screening programs that only collected inspiratory scans become analyzable for small airways disease.

#Impact

LungViT reframes image synthesis as an instrument for an established biomarker: what matters is not the picture but the fSAD percentage computed from it, which tracks lung function decline in two cohorts. The scope is bounded by what the model saw. Development used quality-controlled SPIROMICS protocol scans, and the authors note that low-dose acquisitions, other protocols and contrast-enhanced CT remain untested, while roughly 10% of scans were excluded for a TLC–RV volume difference under 1 L. No code, trained weights, model card or data card have been released, and both cohorts are controlled-access, so applying the method currently means reimplementing it against restricted data.