VERN

Spread through air spaces (STAS) is a lung adenocarcinoma feature defined by where the tumor cells are rather than what they look like: micropapillary clusters, solid nests, or single cells floating in alveolar spaces beyond the edge of the main mass. Because it predicts recurrence after limited resection, it has to be called on an intraoperative frozen section, in the minutes before the surgeon decides between lobectomy and a sublobar resection. Reported accuracy on frozen sections runs 74–85% with sensitivity as low as 55%, and one re-review found 61.3% of STAS cases initially misread as negative.

That geometry is exactly what the standard whole-slide recipe throws away. Multiple instance learning scores each patch on its own and pools the scores, so a cluster of tumor cells reads the same whether it sits inside the tumor body or past its boundary. VERN — a feature-interactive Siamese graph encoder — keeps the arrangement instead. Every 512×512 patch becomes a node, a K-nearest-neighbour graph with K=9 wires each node to its spatial neighbours, and message passing lets the classifier judge a patch in the context of what surrounds it, which is the definition of STAS rather than a proxy.

VERN was built by Hunan University with the departments of pathology and thoracic surgery at the Second Xiangya Hospital of Central South University, released as a preprint in November 2024 and published in npj Precision Oncology the following month.

#Key Features

• Two frozen pathology encoders, one graph: patches are embedded twice — 1024-d by KimiaNet, a DenseNet-121 fine-tuned on TCGA slides, and 768-d by CTransPath, a self-supervised CNN plus Swin Transformer. Both stay fixed; VERN learns only on top.
• Cross-graph weight sharing: the twin encoders are not independent branches. The upper encoder reuses the lower one's weights, so the two representations regularize each other rather than being averaged only at the output.
• Skip connections through the encoder: each branch runs GCNConv → ReLU → SAGEConv → dropout → MLP with the initial node features carried forward, preserving raw patch signal that deep graph message passing otherwise smooths away.
• Readable attention over the slide: per-patch contribution scores are normalized to [0,1] and painted back onto the whole-slide image as a heat map, with the nine highest-contributing patches surfaced for review.
• Frozen and paraffin sections both: training mixes the two, so one checkpoint serves the rushed intraoperative call and the definitive post-operative report.

#Technical Details

The training cohort is 1,546 whole-slide images from 356 lung cancer patients at the Second Xiangya Hospital (206 with STAS, 150 without), each slide labelled by two pathologists working double-blind. Of these, 1,190 slides went to five-fold internal cross-validation and 356 paraffin sections were held back. Tissue is segmented at 20×, tessellated into 512×512 patches, and stain-augmented with a GAN before feature extraction. Training used RMSprop at a learning rate of 0.001, batch size 1 — adjacency matrices differ in size, so graphs cannot be stacked — and 200 epochs on a single RTX 4090. Averaged over the five folds, in-domain testing gave an AUROC of 0.8683 with accuracy 0.7926 and F1 0.7845; the best fold reached 0.9215. Split by section type, AUROC was 0.8829 on paraffin sections and 0.8275 on frozen sections. Multiple instance learning baselines — ABMIL, DSMIL, TransMIL, DTFD-MIL and IBMIL — topped out at 0.7768 AUC on the same data. The trained weights were then applied unchanged to three outside sources: the 356 held-back sections (AUROC 0.9181), 91 slides from the Cancer Hospital of Zhengzhou University and Henan Cancer Hospital (0.8699), 101 TCGA slides (0.7029) and 100 CPTAC slides (0.7555).

#Applications

The intended use is a second read for pathologists calling STAS, most valuably on the intraoperative frozen section that determines resection extent, and the authors run a web platform that accepts an uploaded .svs slide and returns a prediction for centers without the local expertise. The heat maps double as a triage tool, pointing a pathologist at the peritumoral regions where disseminated foci concentrate. The same cohort analysis links STAS to higher PD-L1, P53, Ki67 and ALK expression, tying the prediction to downstream treatment decisions.

#Impact

VERN shows that slide-level spatial topology, not just patch appearance, carries diagnostic signal for an invasion pattern defined by position — and it ships a public checkpoint, a served endpoint, and the extracted features and labels for its external cohorts on Zenodo. The limits are real: it is a narrow binary classifier with no pretraining objective of its own, its accuracy falls off on TCGA and CPTAC because it has seen no data from those centers, it under-detects micropapillary foci, and the raw slides cannot be released under the patient privacy agreement, so reproduction starts from features rather than pixels and needs the KimiaNet and CTransPath weights obtained separately.