Phaet

On the Camelyon cohorts — sentinel lymph-node tiles from breast cancer patients across five Dutch medical centers — the embeddings produced by Phikon-v2 cluster by which center digitized the slide (adjusted Rand index 0.46) and not at all by whether a tile contains metastasis (0.00). The dominant axis of the feature space encodes the scanner, not the tissue. Phaet is that same encoder after a fine-tuning step that inverts the arrangement: the centers become intermixed (ARI 0.01) while tumor and normal tiles separate cleanly (ARI 0.67) — and none of those five centers were seen during fine-tuning, so the invariance generalizes to acquisition sources the model never encountered.

Phaet was released in July 2026 by Waiv, the Paris diagnostics company spun out of Owkin, where the Phikon family originated. It accompanies a preprint by Alexandre Filiot, Oskar Thaeter, Benoît Schmauch and Lionel Guillou applying one model-agnostic fine-tuning recipe to ten pathology foundation models, every one of which improves on both robustness and downstream performance. Two of the ten fine-tuned checkpoints were released: Phaet, from Phikon-v2, and Mascaret, from kaiko.ai's Midnight-12k. Filiot is also first author on H0-mini, an earlier attack on the same problem through distillation.

The proposition is narrow and concrete: Phaet keeps Phikon-v2's architecture, input resolution and 1024-dimensional output, so it swaps into an existing pipeline unchanged, with the invariance baked into the weights rather than bolted on as stain normalization or post-hoc feature correction.

#Key Features

• Invariance instilled in the encoder, not around it: Existing mitigations normalize stain appearance, correct embeddings using site statistics, or suppress site-predictive features during task training — all keeping the backbone frozen. This recipe modifies the backbone itself, producing one task-agnostic robust encoder.
• Largest robustness gain of the ten backbones tested: Phikon-v2 was the least robust model in the study and gains the most, its PathoROB robustness index rising from 0.469 to 0.806, a 72% improvement.
• No accuracy cost: Downstream scores rise alongside robustness on all three evaluation suites, contradicting the expectation that enforcing invariance would erase biologically useful signal.
• Deep adaptation, not a re-tuned output layer: Cross-scanner retrieval measured block by block shows robust matching emerging roughly eight transformer blocks earlier than in the base model, and reaching a higher asymptote.
• Interface-compatible with the model it replaces: A ViT-L/16 taking 224x224 tiles and returning an L2-normalized CLS embedding of dimension 1024, at the same inference cost as Phikon-v2.

#Technical Details

Phaet inherits Phikon-v2's architecture and pretraining: a ViT-L/16 of roughly 303 million parameters trained with DINOv2 on PANCAN-XL, about 456 million H&E tiles drawn from some 60,000 public whole-slide images. Fine-tuning ran on IDRIS resources allocated by GENCI and on EuroHPC's MareNostrum 5 at the Barcelona Supercomputing Center. The preprint characterizes the recipe as lightweight and label-free and reports its effects in detail, but does not specify the procedure; no training code has been released.

Robustness is measured with PathoROB, whose robustness index quantifies how far biological structure dominates medical-center structure in a model's feature neighborhoods across three datasets. Phaet scores 0.785 on TCGA (from 0.619), 0.702 on Camelyon (from 0.019) and 0.932 on Tolkach ESCA (from 0.768), averaging 0.806. Its mean HEST gene-expression Pearson correlation rises from 0.3747 to 0.3943, its THUNDER rank sum falls from 97 to 83 within the paper's 20-model pool, and its Patho-Bench grand average across 63 slide-level tasks rises from 54.1 to 55.8, moving it from 18th to 10th there. Summed across the three suites its rank improves from 57 to 44. These are the model's own figures; the paper's headline averages of 23% robustness and 43% cross-benchmark gain are means over all ten backbones, not Phaet's.

#Applications

Phaet is a frozen tile encoder for computational pathology: embed H&E patches, aggregate with attention-based multiple-instance learning or mean pooling, and fit a light head for biomarker prediction, subtyping, grading, spatial gene expression or survival modeling. It is aimed at cohorts assembled from several laboratories, where scanner and staining differences would otherwise leak into the features and be learned as shortcuts by whatever classifier sits on top.

#Impact

The result reframes acquisition robustness as something repairable after pretraining rather than bought with scale: the recipe acts as an equalizer, closing much of the gap between encoders trained at very different scales. Two limits stand. Phaet at 0.806 remains below the strongest robust encoders — Mascaret reaches 0.924 — so the fine-tune narrows the gap without erasing it, and the robustness index is a feature-space proxy, not a direct measure of downstream performance under domain shift on a clinical task. Weights are gated on HuggingFace behind manual approval under a Waiv non-commercial license limiting use to academic research with attribution, and the model has not been reviewed or approved by any regulatory body for diagnostic use.