Artificial Intelligence Detection of Pulmonary Hypertension
from Lung Transplant Explant Pathology Slides

Background and Rationale

Pulmonary hypertension (mean pulmonary artery pressure, or mPAP, of 20 mmHg or greater at rest, by right heart catheterization) substantially worsens the prognosis of patients with chronic lung disease, including interstitial lung disease (ILD), chronic obstructive pulmonary disease (COPD), and connective tissue disorders. At the time of lung transplantation, the explanted lung provides a uniquely comprehensive tissue sample capturing the full burden of disease across multiple lobes and decades of pathological change.

The blood vessels in these explants show a spectrum of changes that correlate with pulmonary artery pressure: medial (muscular wall) hypertrophy, intimal (inner layer) proliferation and fibrosis, and in severe cases, plexiform lesions -- complex, disorganized vascular growths that represent end-stage vascular remodeling. Prior work by Seki and colleagues (2019), from the Wallace and Saggar research groups at USC and Cedars-Sinai, demonstrated that the presence of capillary proliferation specifically in preserved alveolated lung -- not in fibrotic or destroyed regions -- was the strongest histologic predictor of pulmonary hypertension, with a c-statistic of 0.869. This is an important distinction: the predictive signal is not where one might intuitively look (the obviously diseased regions), but in the relatively preserved parenchyma. Any AI model trained to detect pulmonary hypertension should therefore not restrict itself to grossly abnormal tissue zones.

Separately, Dotan and colleagues (2020) showed that severe histologic vascular changes can be present even when the measured pulmonary artery pressure is not dramatically elevated, suggesting that tissue-based assessment may identify disease severity that hemodynamic measurements alone miss.

Foundation models for digital pathology are large neural networks pretrained on millions of pathology image patches across many tissue types and diseases. They produce rich numerical representations of image patches that capture a wide range of visual features without being trained specifically for any one task. We used phikon-v2, a publicly available pathology foundation model, to extract these representations from every tissue patch across each whole slide, then trained a lightweight classifier on top of those representations to predict pulmonary hypertension at the case level.

Methods

Patient Cohort

Eighteen consecutive lung transplant explant cases from USC Keck Medical Center were included in this analysis, provided by Dr. William D. Wallace. Ten patients had pulmonary hypertension (mean pulmonary artery pressure 26 to 51 mmHg, measured by right heart catheterization at the time of transplant listing). Eight patients had no pulmonary hypertension and served as controls. Slides were digitized at 20x magnification using an Aperio scanner, producing 128 whole-slide images in total. The number of slides per patient ranged from 3 to 10 (median 7). Some slides contain expert annotations of anatomical compartments; these annotations were not used in this analysis and are reserved for a planned follow-up analysis.

Slide Quality Control

All 128 slides were screened for pen/grossing ink, tissue fold, focus artifact, and dark spot using an in-house gated-attention multiple instance learning quality control classifier (Track B; pen AUROC 0.813 on an independent BRAF cohort). No pen or grossing ink artifact was identified in the WDW cohort. All slides were retained for analysis.

Image Processing and Feature Extraction

Each whole-slide image was divided into non-overlapping 256x256-pixel square patches at 20x magnification. Patches containing predominantly background (glass or white space) were excluded using tissue masking. The remaining tissue patches were each processed through phikon-v2, a pathology foundation model based on a Vision Transformer Large (ViT-L) architecture that produces a 1,024-dimensional numerical feature vector for each patch. These vectors were stored for each slide along with the spatial coordinates of each patch within the slide. Feature extraction was performed on an NVIDIA RTX 5090 graphics processing unit and required one overnight run to process all 128 slides.

Classifier Architecture

We used attention-based multiple instance learning (ABMIL), a framework designed for whole-slide image analysis in which the final prediction is made at the case (patient) level rather than at the patch level. The model learns to assign an importance weight (attention score) to each image patch, then combines all patches from a patient's slides into a single case-level prediction using those weights as a weighted average. Patches the model finds informative receive high attention scores; uninformative patches are down-weighted. For each patient, all patches from all available slides were combined into one bag of feature vectors fed to the model together.

The classifier itself consists of: (1) a patch encoder that projects the 1,024-dimensional features to a 512-dimensional space using a linear layer, layer normalization, and a nonlinear activation; (2) a gated attention mechanism using parallel tanh and sigmoid pathways that produces a scalar attention weight for each patch; (3) a weighted aggregation step that produces one 512-dimensional vector per patient; and (4) a two-layer classification head that produces the final probability of pulmonary hypertension. The model was trained using the AdamW optimizer, a cosine learning rate schedule, and binary cross-entropy loss. Early stopping was applied when training loss did not improve for 20 consecutive training epochs, with a maximum of 60 epochs. During training, patches were randomly subsampled to a maximum of 8,192 per patient to manage memory; at test time, all patches were used.

Evaluation: Leave-One-Out Cross-Validation

With only 18 patients, standard multi-fold cross-validation would leave too few patients in each test set to provide meaningful evaluation. We therefore used leave-one-out cross-validation: for each patient, a fresh model was trained from scratch on all 17 other patients, and that model was then applied to the held-out patient whose slides it had never seen. This produced one predicted probability of pulmonary hypertension for each of the 18 patients. These 18 probabilities were then used to compute the area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, and classification accuracy. The Youden index was used to identify the probability threshold that maximized the sum of sensitivity and specificity. Total runtime: 77 minutes on an RTX 5090 graphics processing unit.

Results

Receiver Operating Characteristic Curve

Per-Patient Results

The model produced highly confident predictions for most cases. Fifteen of 18 patients received predicted probabilities below 0.005 or above 0.999, indicating strong separation between the two groups. The two cases the model missed have clear explanations (see Discussion).

Discussion

This study provides the first demonstration that a foundation model artificial intelligence system can detect pulmonary hypertension from lung transplant explant whole-slide images with clinically meaningful accuracy, using only slide images and the clinical hemodynamic diagnosis -- no region-by-region annotation by a pathologist was required.

Confidence of Predictions

The striking feature of the results is not just the overall accuracy but the extreme confidence of most individual predictions. Probabilities were clustered near 0 (control-like) or near 1 (PH-like), with very little in between. This bimodal distribution suggests the model has learned a global texture or architectural signature that differs qualitatively between pulmonary hypertension and non-PH explants, rather than detecting rare focal lesions. This is consistent with the histological understanding of PH, in which vascular remodeling is a diffuse, multifocal process involving many vessels throughout the lung.

The perfect specificity (zero false positives) is particularly noteworthy from a clinical standpoint. A screening tool that produces no false alarms -- where every case it calls pulmonary hypertension actually has it -- is highly practical. This likely reflects the distinct and consistent global appearance of PH tissue versus non-PH tissue even in the setting of other underlying lung disease.

The Two Missed Cases

Relationship to the Seki et al. (2019) Findings

Seki and colleagues demonstrated that the pathologist-assessed presence of capillary proliferation in preserved (non-fibrotic) alveolated lung had a c-statistic of 0.869 for predicting pulmonary hypertension -- a strong result from careful manual assessment. The Computational Pathology Kit (CPK), requiring no manual assessment at all, achieved an AUROC of 0.975 in a leave-one-out analysis. The attention maps shown in the Attention Maps tabs of this report provide visual evidence about which tissue regions drove the AI predictions, and may confirm or extend the finding that preserved parenchyma (rather than fibrotic or destroyed zones) is where the discriminative signal lives.

Limitations

The cohort size of 18 cases is small. Leave-one-out cross-validation provides an unbiased estimate but the resulting confidence intervals are wide. All cases are from a single institution, a single scanner, and a single pathology processing protocol. The pulmonary hypertension etiology varies across cases (Group 3 pulmonary hypertension associated with lung disease being the most common in this transplant population) and this variation has not been analyzed separately. The model has not been tested on cases from any other institution. Prospective external validation is essential before any clinical conclusions can be drawn.

Validation Cohort Strategy

No publicly available labeled pulmonary hypertension lung transplant explant whole-slide image dataset exists (as of May 2026). The most realistic path to external validation is via expansion of the USC/Keck cohort or collaboration with other high-volume lung transplant programs. A multi-institutional dataset of 50–90 cases would provide the sample size needed for publication-grade validation.

Conclusions

Attention-based multiple instance learning applied to phikon-v2 foundation model features achieves an AUROC of 0.975, sensitivity of 90%, specificity of 100%, and zero false positives for detecting pulmonary hypertension from lung transplant explant whole-slide images in leave-one-out cross-validation across 18 cases. The two misclassifications have straightforward clinico-pathologic explanations: the mildest disease in the cohort, and the case with the fewest available slides. The system requires no pathologist annotation and can process a full case in minutes once features are extracted. External validation in a multi-center cohort is the clear next step.

References

1. Seki N, et al. Histopathologic features predictive of pulmonary hypertension in patients with interstitial lung disease. Chest. 2019. [Wallace and Saggar groups, USC / Cedars-Sinai]. Capillary proliferation in preserved lung: c-statistic 0.869.
2. Dotan Y, et al. Pulmonary hypertension in patients with interstitial lung disease -- relationship between histological vasculopathy and hemodynamics. Chest. 2020. Severe vasculopathy independent of mean pulmonary artery pressure in ILD explants.
3. Filiot A, et al. phikon-v2: A Vision Foundation Model for Pathology. HuggingFace Hub. 2024. Vision Transformer Large architecture, 1,024-dimensional features, trained on over 460,000 whole-slide images.
4. Ilse M, Tomczak JM, Welling M. Attention-based Deep Multiple Instance Learning. International Conference on Machine Learning (ICML). 2018. Foundational paper for the gated attention aggregation method used in this work.
5. Simonneau G, et al. Updated clinical classification of pulmonary hypertension. Journal of the American College of Cardiology. 2019. Mean pulmonary artery pressure threshold revised to 20 mmHg.
6. Humbert M, et al. 2022 ESC/ERS Guidelines for the diagnosis and treatment of pulmonary hypertension. European Respiratory Journal. 2022.