HATS

A lymph node biopsy arrives at a hematopathology service as an H&E-stained slide, but the H&E almost never settles the diagnosis. A pathologist must preview it, decide which immunohistochemistry (IHC) stains to order, wait for those stains, then review the case again. The preview takes minutes; the coordination around it — slide distribution, order entry, communication across the clinical team — adds a day of turnaround, and more than 90% of hematopathology specimens require it.

HATS, the Hematopathology Automatic Triaging System, moves that ordering decision to the moment the slide is scanned. It cuts the whole-slide image into non-overlapping patches at 20x, encodes each with a frozen pathology foundation model, and aggregates them into a slide-level probability distribution over ten lymphoma categories with a gated attention-based multiple-instance-learning network. The top prediction is mapped onto one of three institutional IHC panels — high-grade B-cell, low-grade B-cell, or disease-specific — which is ordered automatically. The design leans on what makes triage an unusually forgiving target for automation: an unnecessary stain is caught at pathologist review before it can influence any clinical decision. Grouping related subtypes behind a shared panel compounds that tolerance, since a case confused between diffuse large B-cell and follicular lymphoma still receives the correct high-grade B-cell panel.

HATS was built at Memorial Sloan Kettering Cancer Center and posted to medRxiv in August 2026. A systematic comparison of seven public pathology foundation models on this task preceded it; H-optimus-0 emerged strongest and became the encoder used throughout.

#Key Features

• Frozen backbone, trained head: Foundation model weights never update, so only the attention classifier is fit. Two 128-dimensional gated projections and a 1536-to-512-to-10 head are the whole trainable model, which made a seven-encoder sweep affordable.
• Panel-level output: Predictions become institutional IHC panel orders rather than raw subtype calls, aligning output with the decision a hematopathologist actually makes.
• Case-level aggregation: Slides from the same biopsy are combined by entropy-weighted averaging, lifting accuracy from 79.4% slide-level to 83.8% case-level.
• Distributed attention: The three highest-weighted patches receive only 5.3% of total attention, and deleting the top five costs 0.33% accuracy: predictions rest on diffuse morphology, not signature tiles.
• Calibrated confidence: Only 8% of correct predictions fall below 0.6 confidence, against more than 42% of incorrect ones, so a deferral threshold cuts erroneous orders by over 40%.

#Technical Details

Slides were digitized at 40x, preprocessed with the Trident pathology library, and tissue-masked with the GrandQC artifact segmentation network before 20x patch extraction. The gated attention module computes instance scores from tanh- and sigmoid-activated projections of the 1,536-dimensional H-optimus-0 patch features, and the attention-weighted sum feeds a two-layer classifier under 0.4 feature dropout. The cohort comprised 4,996 whole-slide images from 1,607 patients, split 70/15/15 at the patient level. On the held-out test set the model reaches 0.962 ± 0.009 macro ROC-AUC, 83.8% ± 1.5% case-level accuracy, and 92.0% ± 1.8% IHC panel accuracy. Across encoders, case-level accuracy runs H-optimus-0 83.8%, Virchow2 80.0%, UNI v2 79.4%, Prov-GigaPath 78.3%, CONCH 73.9%, MUSK 72.7%, and Phikon v2 71.1%; concatenating top performers' features yielded no gain. A parameter-free k-nearest-neighbor classifier over mean-pooled H-optimus-0 embeddings already reaches 68% slide-level accuracy, which the attention mechanism raises to 79%.

#Applications

In a blinded reader study on the held-out set, HATS classified 84.7% of cases correctly against 65.3% for practicing pathologists restricted to morphology alone across 137 evaluations — a comparison of the triage step, not of diagnosis, since pathologists normally have clinical history and flow cytometry in hand. It was then run on an unfiltered consecutive series of 230 cases and 647 whole-slide images from the MSKCC hematopathology service over five months. Seven scant-tissue cases went to manual review; of the remaining 223, an expert pathologist judged the ordered panels fully sufficient for diagnosis in 72.6% of cases, partially sufficient in 18.4%, and wrong in 9%.

#Impact

HATS demonstrates that a lightweight classifier on a public, general-purpose pathology encoder reaches clinically useful performance without institution-scale pretraining — a path open to any service with a slide archive and modest compute. Its limits are equally clear. The label space is closed at ten categories, and only 59% of the consecutive clinical cases fell inside it; accuracy for angioimmunoblastic T-cell, anaplastic large cell, and Burkitt lymphoma falls below 60%, and roughly half the mis-triaged cases were T-cell lymphomas. Both cohorts come from one institution with its own panel conventions, and the slides cannot be shared under MSKCC patient privacy policy. The checkpoint repository named in the preprint is public under Apache 2.0 but holds no weight files, and the GitHub repository named in the code availability statement is not public. It remains a preprint awaiting peer review, and the authors are explicit that the final diagnosis stays with the pathologist.