GASPath

Under the microscope, a well-differentiated endocervical gastric-type adenocarcinoma (GAS) can be almost indistinguishable from a benign endocervical gland: bland nuclei, few mitoses, abundant clear or eosinophilic cytoplasm, and a paradoxical "well differentiated but deeply invasive" growth pattern. The consequence is not academic. GAS is HPV-independent, its 5-year survival is roughly 42% against 91% for HPV-associated adenocarcinoma, and 59% of cases are already at FIGO stage II or later when they are found. Confirmatory immunohistochemistry, in situ hybridization and sequencing are expensive, slow, and often unavailable.

GASPath answers that with a slide-level classifier that reads routine H&E alone. It is a multiple-instance-learning model: the whole-slide image is tiled into patches, each patch is embedded once by a frozen pathology foundation model, and a trained aggregator collapses the resulting bag of embeddings into one diagnosis. The encoder is Virchow2, used off the shelf and never retrained on the cervical cohort; GASPath's own contribution is the aggregation stage. Rather than let every patch attend to every other patch, a small set of residual "agents" stands in for the queries, keys and values, so attention propagates through a lower-dimensional space, and a soft-mask filter suppresses redundant local features before pooling. The work came from the Department of Pathology at Peking University Third Hospital together with the Institute of Biopharmaceutical and Health Engineering at Tsinghua University's Shenzhen International Graduate School.

#Key Features

• Agent-aware attention aggregation: Residual agents replace queries, keys and values as the carriers of attention, and a soft-mask filter strips redundant local signal, sharpening slide-level pooling relative to standard self-attention.
• Frozen foundation-model encoder: Patch embeddings come from a pathology foundation model that is not extend-pretrained here; the only fitted parameters are the aggregator and decision head.
• A five-way differential, not a binary alarm: Output spans normal mucosa, benign endocervical lesions (tunnel clusters and microglandular hyperplasia), GAS, HPV-associated adenocarcinoma, and endometrioid carcinoma with mucinous differentiation — the entities GAS is actually confused with.
• Operating points fixed before external use: Two GAS-probability thresholds, 0.20 and 0.06, were set on internal validation and held constant across every external cohort, trading a little specificity for sensitivity.
• Interpretable output: Each prediction carries an attention heatmap overlaid on the slide plus the ten highest-attention patches, which localize irregular angulated glands, desmoplastic stroma, and loss of nuclear polarity.

#Technical Details

Slides are digitized at 0.105 µm/pixel and tiled at a fixed 0.84 µm/pixel with 224 × 224 patches, near 10× magnification, which ablation showed beats 2.5×. Training drew 1,629 slides from Peking University Third Hospital — 309 GAS slides from 96 HPV-negative, surgically confirmed cases plus 1,320 slides across the four comparison categories — with multiple slides per case taken from distinct paraffin blocks to prevent leakage; the aggregator was fit by 10-fold cross-validation on a 6:2:2 split. Internal accuracy was 0.980 (95% CI 0.977–0.983), F1 0.935, ROC-AUC 0.995 (0.994–0.997), and Cohen's κ 0.926 on the five-way task. One checkpoint then transferred unchanged to 249 GAS slides from 12 external centres, reaching 0.902 sensitivity under argmax and 0.968 at the relaxed threshold, and to 7,056 unselected consultation slides collected between March 2024 and April 2025, where balanced accuracy was 0.953 and all 45 GAS cases were identified. On biopsy specimens, where shallow sampling is the usual failure mode, ROC-AUC was 0.990 (0.984–0.997). Encoder ablation against CONCH, PLIP and an ImageNet-pretrained ResNet, and aggregator comparison against CLAM, TransMIL and ABMIL, support both design choices.

#Applications

The intended use is triage in routine gynecological pathology: flagging cervical specimens that deserve a second read or confirmatory immunohistochemistry and in situ hybridization, not replacing them. GAS made up 0.6% of the real-world cohort, so the operating regime is high-sensitivity screening against a large benign background — the situation that matters most for colposcopy-guided biopsies and for laboratories without ready access to molecular assays. A positive call can also prompt ancillary testing for HER2 amplification or Claudin 18.2 expression, both of which carry therapeutic consequences.

#Impact

GASPath is narrow by design: one closed-set diagnostic task, no pretraining objective of its own, and its interest lies in the validation discipline rather than the architecture. A single checkpoint with pre-set thresholds was carried across 13 Chinese centres and a consecutive year of consultation material without per-cohort refitting, which is the evidence standard clinical deployment requires. The authors state the limits plainly: the cohorts were assembled retrospectively and some are consultation-based, biasing toward hard cases; all centres are in China, so generalization across populations, ethnicities and scanning protocols is unverified; no domain-adaptation technique was applied; and the large-scale stage was scored offline rather than inside a live diagnostic workflow. The training and inference code is released under an MIT licence, but trained weights are not distributed and are offered on request.