HF Lung V1: High-Fidelity Lung Models

• HF Lung V1 is a first-generation high-fidelity lung modeling framework that reconstructs pulmonary structure, mechanics, and function using clinically acquired data.
• It employs modular pipelines across modalities—ventilator waveforms, imaging segmentation, MRI, and CFD—to capture physiologic features with interpretable, parameter-driven designs.
• The approach balances physiologic accuracy with observability, enabling actionable insights for ARDS, VILI, and other pulmonary conditions while remaining estimable from routine data.

“HF Lung V1” is usefully understood as an Editor’s term for a first-generation class of high-fidelity lung models and pipelines that reconstruct pulmonary structure, mechanics, or function from clinically acquired data. In the literature assembled here, that label spans a “damage-informed” ventilator waveform model built directly on pressure–volume traces, deep segmentation systems for chest radiography and CT, multiscale airflow and cardiopulmonary circulation models, non-contrast functional MRI pipelines, X-ray velocimetry for regional ventilation, CT-radiomics functional surrogates, and lung-ultrasound prognostic models (Agrawal et al., 2020, Pusterla et al., 2024, Bauman et al., 17 Apr 2026). The common thread is not a single software package but a recurring design objective: preserve physiologic or anatomic interpretability while remaining estimable from routine or near-routine data.

1. Conceptual scope and nomenclature

The most explicit “V1” framing appears in the discussion of the ventilator-waveform model for acute respiratory distress syndrome and ventilator-induced lung injury. That work states that the paper does not explicitly label the model as “V1,” but conceptually treats it as the first version of a damage-informed, waveform-based high-fidelity lung model that reconstructs pressure and volume signals from physiologically interpretable subcomponents and estimates parameters from ventilator data (Agrawal et al., 2020). In that source, “high-fidelity” does not mean a fully anatomical alveolar simulation; it means a model that sits between the single-compartment equation and highly detailed multi-compartment or spatial models by modeling pressure and volume waveforms explicitly in time, encoding physiologic features as modular parameters, and remaining estimable from routine ventilator data.

The broader literature extends the same first-generation, high-fidelity logic to other modalities. In chest radiography, Deep LF-Net defines a robust end-to-end DeepLabv3+ segmentation pipeline with encoder–decoder structure, atrous convolution, and ASPP, designed to handle severe abnormal findings without pre-processing (Singh et al., 2020). In CT, a multi-task V-Net uses simultaneous lobar and tracheobronchial segmentation to improve lobar delineation in diseased lungs (Martell et al., 2021). In functional MRI, TrueLung provides an automated path from free-breathing ultra-fast balanced SSFP acquisitions to ventilation and perfusion maps, defect metrics, and concise reports (Pusterla et al., 2024). In non-contrast-enhanced MRI under physiological non-stationarity, VQ-Wave replaces spectral decomposition with a physics-driven spatio-temporal neural network (Bauman et al., 17 Apr 2026). This suggests that “HF Lung V1” is best treated as a family resemblance concept rather than a standardized product name.

A second unifying feature is modularity. Across the cited work, modules recur in different forms: waveform submodels for inspiratory and expiratory slopes, auxiliary segmentation heads for bronchi and lobes, matrix-pencil decomposition blocks for respiratory and cardiac oscillations, and 3D–1D coupling between resolved airways and distal trees. A plausible implication is that V1 systems privilege decomposability and observability over maximal physiologic completeness.

2. Physiologic and cardiopulmonary modeling

The canonical mechanics-oriented HF Lung V1 formulation is the ventilator-waveform model for ARDS and VILI. It is described as a modular, lumped waveform model in which pressure and volume are each represented as a sum of submodels rather than by the classic single-compartment governing equation $P(t)=R\dot{V}(t)+EV(t)+P_0$ (Agrawal et al., 2020). The volume model starts from a sinusoid $f_{s_1}(t)=\sin(2\pi \theta t-\phi_1)-b_1$, converts it into a rectangular inspiratory–expiratory base with a hyperbolic tangent, and constructs inspiratory and expiratory recursions governed by $\beta_1$ and $\beta_2$, with final volume $V(t)=A_v(f_{v_1}(t)+f_{v_2}(t))$. The pressure model likewise uses periodic bases and combines an inspiratory/expiratory rate submodel, a plateau-peaks submodel, a low-volume inspiratory submodel, and baseline pressure $A_{p_4}$ as PEEP. Its parameters are explicitly tied to clinically used waveform features such as inspiratory and expiratory gradients, plateau peaks, and low-volume recruitment behavior.

This parameterization is deliberately physiologic. Larger $\beta_1$ indicates slower rise in volume and can reflect lower inspiratory flow or lower compliance; larger $\beta_2$ indicates slower expiratory fall and relates to a longer expiratory time constant; larger $A_{p_1}$ indicates higher plateau pressure for the same volume and therefore increased elastance; $A_v/A_{p_1}$ functions as a compliance proxy; and the plateau perturbation parameters $f_{s_1}(t)=\sin(2\pi \theta t-\phi_1)-b_1$0 quantify inspiratory resistance, patient effort, or dyssynchrony when those waveform features are present (Agrawal et al., 2020). The estimation strategy is equally central to the model’s “V1” status: only parameters associated with features present in a given breath are estimated, low-impact parameters are fixed, parameter bounds are chosen iteratively to avoid non-physiologic behavior, and uncertainty is assessed with 1000 constrained optimization runs using MATLAB fmincon with MultiStart.

A related cardiopulmonary extension appears in the closed-loop lumped-parameter model of cardiovascular function after lung resection. There, each chamber obeys the time-varying elastance law $f_{s_1}(t)=\sin(2\pi \theta t-\phi_1)-b_1$1, valves are represented by nonlinear resistance and inertance, and pulmonary and systemic circulations are each modeled as three-element Windkessels (Huang et al., 2 May 2025). Lung resection is encoded as a reduction in the number of pulmonary vascular segments, producing

$f_{s_1}(t)=\sin(2\pi \theta t-\phi_1)-b_1$2

with resection fraction $f_{s_1}(t)=\sin(2\pi \theta t-\phi_1)-b_1$3. RV contractility loss is represented by reducing $f_{s_1}(t)=\sin(2\pi \theta t-\phi_1)-b_1$4. A major mechanistic result is that afterload increase and contractility loss generate similar reductions in RVEF and similar RV dilation, but opposite changes in RVSP, PASP, and PADP: afterload increase raises them, whereas contractility loss lowers them (Huang et al., 2 May 2025). That distinction provides an interpretable axis for right-heart failure phenotyping in lung-resection and pulmonary vascular disease settings.

3. Anatomical reconstruction, segmentation, and volumetry

In imaging-centric HF Lung V1 implementations, segmentation is the foundational operation. For frontal chest radiographs, Deep LF-Net uses DeepLabv3+ with encoder–decoder structure, atrous convolution, and ASPP, with atrous rates $f_{s_1}(t)=\sin(2\pi \theta t-\phi_1)-b_1$5, an output stride of 16, and pre-trained ResNet18 or MobileNetv2 backbones (Singh et al., 2020). All images are resized to $f_{s_1}(t)=\sin(2\pi \theta t-\phi_1)-b_1$6, and the paper explicitly states that no pre-processing technique is required before the image is fed to the network. False positives are removed by morphological area filtering that keeps only the two largest connected components. On the Indian dataset, which includes 688 PA chest radiographs with severe abnormalities, the reported results are 99.52% accuracy, 98.95% sensitivity, 99.71% specificity, 95.21% Jaccard, and 95.28% Dice with ResNet18, and 99.49% accuracy, 98.57% sensitivity, 99.79% specificity, 98.16% Jaccard, and 97.20% Dice with MobileNetv2 (Singh et al., 2020).

For thoracic CT, the multi-task V-Net formulation addresses the specific problem of lobar segmentation in distorted anatomy. It accepts $f_{s_1}(t)=\sin(2\pi \theta t-\phi_1)-b_1$7 volumes, uses a 3D V-Net encoder–decoder with residual connections, PReLU, batch normalization, dropout, and attention-gated skip pathways, and branches into a main lobar head and an auxiliary tracheobronchial head with a multi-task Dice-style objective $f_{s_1}(t)=\sin(2\pi \theta t-\phi_1)-b_1$8, with $f_{s_1}(t)=\sin(2\pi \theta t-\phi_1)-b_1$9 (Martell et al., 2021). External validation on six-case cohorts with COPD, lung cancer, COVID-19 pneumonitis, and collapsed lung produced mean per-segment Dice scores of 0.94, 0.94, 0.94, and 0.92, respectively, while normal lungs were reported at 0.97 (Martell et al., 2021). The explicit rationale is that airway topology constrains lobar extent when fissures are incomplete or obscured.

Pulmonary vascular segmentation forms a parallel anatomic branch. The contrast-enhanced CT pipeline based on an offset medialness function first extracts the airway tree by iterative region growing, then separates left and right lungs, constrains vessel enhancement to lung masks, and reconstructs a pulmonary vascular tree with centerline extraction, Dijkstra-like reconnection, radius estimation, and total-variation geodesic active contour refinement (Helmberger et al., 2013). Its downstream quantitative target is pulmonary hypertension phenotyping through vessel tortuosity. The reported Spearman correlation between distance-metric tortuosity and mean pulmonary artery pressure is $\beta_1$0 with $\beta_1$1, and the group comparison between PH and non-PH is significant at $\beta_1$2 (Helmberger et al., 2013).

Volumetry from projection imaging extends the same logic from anatomy to global lung size. A U-Net trained on synthetic radiographs and CT-derived pixel-level lung thickness maps estimates total lung volume from frontal chest radiographs (Dorosti et al., 2021). The thickness map represents the local path length of lung tissue along the X-ray beam, and TLV follows by summation over lung pixels. On real KRI radiographs, the reported performance after PA-diameter correction is $\beta_1$3, $\beta_1$4, and $\beta_1$5 against CT-derived reference TLV (Dorosti et al., 2021). This introduces a projection-based volumetric branch that is lower fidelity than CT but much more deployable.

A malignancy-oriented extension appears in AutoRad-Lung, which couples a pre-trained AIMv2 vision encoder with prompts generated from hand-crafted radiomics through conditional context optimization (Khademi et al., 26 Mar 2025). The feature vector $\beta_1$6 is mapped by a Meta-Net to a prompt perturbation $\beta_1$7, which modifies learnable context tokens before class text embeddings are encoded. In the three-class benign–malignant–unsure formulation on LIDC-IDRI, the reported accuracy rises from $\beta_1$8 for CLIP-Lung to $\beta_1$9 for AutoRad-Lung, with marked improvement in the unsure class (Khademi et al., 26 Mar 2025). Although this is not a lung mechanics model, it exemplifies a V1 design in which quantitative imaging descriptors directly condition multimodal inference.

4. Functional ventilation and perfusion mapping

The richest functional HF Lung V1 pipelines are MRI-based. TrueLung starts from free-breathing, contrast-agent-free, time-resolved ultra-fast balanced SSFP MRI, applies image quality checks, Graph Diffusion Regularization registration, MD-GRU whole-lung and lobar segmentation, and matrix pencil decomposition to produce voxel-wise fractional ventilation $\beta_2$0, perfusion $\beta_2$1, and blood-arrival-time maps (Pusterla et al., 2024). Defect quantification uses thresholds at 75% of the voxelwise median within a mask: $\beta_2$2 with RFV and Ro/RQ computed as the percentage of voxels below threshold, area-weighted across slices. In 75 children with cystic fibrosis, the whole pipeline required about 20 minutes per subject, automated whole-lung quantification was satisfactory in 88% of patients and 97% of slices, automated lobar quantification in 73% of patients and 93% of slices, and differences between fully automated and manually refined RFV/RQ were described as marginal (Pusterla et al., 2024).

VQ-Wave addresses a specific failure mode of spectral decomposition: physiological non-stationarity. Its voxel signal model is

$\beta_2$3

with amplitude- and frequency-modulated oscillatory ventilation and perfusion components trained synthetically across amplitude drift, frequency drift, sighs, and noise (Bauman et al., 17 Apr 2026). The network is a spatio-temporal inception-style CNN with squeeze-and-excitation attention, 3×3 spatial patch context, multi-scale temporal kernels $\beta_2$4, and an eight-parameter regression head for amplitudes, frequencies, and phase encodings. In numerical and in-vivo comparisons against matrix pencil decomposition, VQ-Wave maintained mean variation below 12% when scan time was reduced from 45 s to 15 s, whereas MP showed severe degradation under irregular physiology and short acquisitions (Bauman et al., 17 Apr 2026). This directly generalizes the TrueLung design philosophy from deterministic decomposition to learned but physics-driven decomposition.

Regional ventilation can also be obtained from X-ray motion. X-ray Velocimetry uses five fluoroscopic projection angles around the thorax together with a breath-hold CT to reconstruct local tissue motion and derive voxelwise specific ventilation, $\beta_2$5, in approximately 5000 voxels across the lung (Smith et al., 27 Feb 2025). Region-level metrics are mean specific ventilation,

$\beta_2$6

and ventilation heterogeneity,

$\beta_2$7

In the sheep endobronchial-valve study, XV visualized and quantified reduction of airflow downstream of the valves both where collapse was and was not visible in CT, and also showed compensatory changes in non-target lung regions (Smith et al., 27 Feb 2025). The practical significance is that function can be perturbed before structural collapse becomes evident.

Static CT textures offer another route to function. In the 4DCT plus DTPA-SPECT study, 79 radiomic features were screened patch-wise, eight non-redundant features showed medium-to-large effect size for distinguishing defected from non-defected patches, and voxel-wise feature maps were then generated with a $\beta_2$8 sliding window (Huang et al., 2022). Among the candidates, the phase-averaged feature map of GLDM Dependence Non-uniformity achieved median SCC 0.60 against SPECT ventilation, median DSC 0.60/0.65 for high/low functional lung partitions, ICC 0.96 across ten phases, and correlation 0.646 between spatially averaged feature values and PFT measurements (Huang et al., 2022). This is important because it decouples functional inference from deformable registration and uses a static-texture surrogate with explicit radiomic semantics.

Lung ultrasound brings the same functional emphasis to heart failure. In the CHF readmission pilot, spatiotemporal embeddings from a pretrained TSM ResNet-18 were extracted from six standardized B-mode lung views, and temporal differences between sequential examinations were more predictive than single-timepoint representations (Armouti et al., 16 May 2026). The best multi-view MLP reached an F1 score of 0.80 with 95% CI 0.62–0.96; the dependent lower-lung views Left-3 and Right-3 carried the strongest single-view signal; and pleural-line abnormalities, including breaks and indentations, were reported to be as informative as canonical A-line and B-line markers (Armouti et al., 16 May 2026). This extends HF Lung V1 beyond pulmonary parenchymal mapping to prognostic monitoring of congestion.

5. Multiscale airflow, flow–volume behavior, and subject-specific CFD

A separate branch of HF Lung V1 is explicitly multiscale and physics-based. One formulation couples a hierarchical bronchial tree with a one-dimensional elastic parenchyma representation and solves the resulting linear air–tissue interaction problem in the frequency domain (Dio et al., 2022). The airway tree has four levels: a lumped nasopharyngeal resistance, explicit rigid cylindrical proximal bronchi, symmetric equivalent-resistance intermediate airways, and acinar trees; the distal pressure–flow relation is written $\beta_2$9, with $V(t)=A_v(f_{v_1}(t)+f_{v_2}(t))$0 the aggregated resistance matrix (Dio et al., 2022). Regional flows obey $V(t)=A_v(f_{v_1}(t)+f_{v_2}(t))$1, total flow is the sum of regional flows, and flow–volume curves are reconstructed by integrating the resulting mouth flow. The major physiological conclusion is that obstruction alters expiratory curve shape much more strongly than inspiration, while compliance changes have more limited influence on the end of expiration and the start of inspiration (Dio et al., 2022). The use of the Discrete Fourier Transform enables quasi-instantaneous solving under the model hypotheses.

A higher-resolution counterpart is the multiscale CFD breathing-human-lung model built from subject-specific MDCT at total lung capacity and functional residual capacity (Choi et al., 2013). Its proximal 3D geometry extends from mouth to about seven airway generations, while an anatomically consistent 1D tree spans more than twenty generations to terminal bronchioles. A registration-derived displacement field between TLC and FRC specifies regional ventilation and deforms the airway surface mesh, thereby providing subject-specific flow boundary conditions for a breathing simulation (Choi et al., 2013). Large eddy simulation then resolves inhalation and exhalation flow structures. The reported phenomena include a turbulent laryngeal jet on inhalation, oscillatory jets and elongated vortical tubes on exhalation, and spatial hot spots of wall shear stress that are proposed as sites for future mechanotransduction investigation (Choi et al., 2013). If the waveform model and the DFT model represent clinically estimable V1 abstractions, this CFD framework represents the same V1 ambition at the subject-specific mechanistic end of the spectrum.

The relation between these two multiscale formulations is instructive. The DFT model emphasizes identifiability and rapid parameter sweeps; the 3D+1D LES model emphasizes geometric realism and resolved proximal turbulence. This suggests that HF Lung V1, in its most inclusive sense, spans both ends of a fidelity–estimability continuum, provided the modeling choices remain anchored to measurable outputs such as flow–volume curves, pressure drops, or wall shear stress.

6. Validation domains, limitations, and likely evolution

Across the cited literature, HF Lung V1 appears in several validation domains: ARDS/VILI waveform phenotyping, pulmonary lobar segmentation, lung nodule malignancy classification, cystic-fibrosis functional MRI, emphysema intervention monitoring, pulmonary-hypertension vascular analysis, cardiopulmonary consequences of lung resection, and CHF readmission prediction (Agrawal et al., 2020, Martell et al., 2021, Pusterla et al., 2024, Smith et al., 27 Feb 2025, Huang et al., 2 May 2025, Armouti et al., 16 May 2026). The common validation pattern is multimodal comparison against an independent reference: waveform parameters against PV loops and single-compartment analysis, MRI defect maps against manually refined segmentations, CT feature maps against SPECT ventilation, vascular tortuosity against right-heart-catheterization-derived mPAP, and CXR-derived TLV against CT reference volume.

The limitations are equally recurrent. The waveform model avoids the “too complex to estimate” problem of multi-compartment lung models but does not explicitly model gas exchange or vascular permeability (Agrawal et al., 2020). Deep LF-Net and the multi-task V-Net generalize to severe abnormalities, yet both papers still highlight failures in extreme pathology and the need for more such cases in training (Singh et al., 2020, Martell et al., 2021). The multiscale flow models assume rigid airways, linear elasticity, or symmetry in distal trees, and are less suited to strongly nonlinear or severely diseased states (Dio et al., 2022, Choi et al., 2013). TrueLung is off-line and was evaluated in a pediatric CF cohort from a single vendor, while VQ-Wave, although robust to irregular physiology, remains a pilot in-vivo study with six human subjects (Pusterla et al., 2024, Bauman et al., 17 Apr 2026). The CHF ultrasound model is explicitly a pilot with 30 patients and wide confidence intervals (Armouti et al., 16 May 2026). The CT-radiomics functional surrogate study used 21 patients, with PFTs available in only 13 (Huang et al., 2022). These are not incidental caveats; they define the empirical perimeter of a V1 generation.

The literature also sketches an evolution path. The waveform paper proposes tighter coupling of pressure and volume, possible gas-exchange or perfusion integration, and automated phenotype classification (Agrawal et al., 2020). TrueLung points to federated learning, attention-based segmentation improvements, and additional defect metrics such as V/Q overlap and defect distribution index (Pusterla et al., 2024). VQ-Wave suggests larger clinical trials, more realistic simulators, and adaptation across field strengths and organs (Bauman et al., 17 Apr 2026). The lung-resection circulation paper identifies ventricular interdependence, spatial pulmonary branches, and regulatory feedback as next steps beyond the initial two-Windkessel architecture (Huang et al., 2 May 2025). This suggests that “V1” is best read literally: a first, modular, clinically tractable layer upon which tighter heart–lung coupling, richer gas transport, and more standardized multi-center validation can be added.

Taken together, HF Lung V1 denotes a first-generation attempt to make lung structure and function computationally explicit at clinically useful resolution. In one branch it means waveform-based inference of compliance, resistance, recruitment, and dyssynchrony from ventilator traces. In another it means automated segmentation of lungs, lobes, vessels, or nodules from radiographs and CT. In another it means regional ventilation and perfusion mapping from MRI, fluoroscopy, or CT textures. In yet another it means multiscale airway and parenchymal simulation or closed-loop heart–lung hemodynamics. The cumulative record does not define a single canonical model, but it does define a coherent research program: modular, physiologically anchored, high-fidelity lung inference that remains estimable from real clinical data.