Phikon

Phikon is a family of self-supervised vision foundation models for computational pathology developed by Owkin, the French-American AI biotech company. The models learn general-purpose tile-level representations from hematoxylin and eosin (H&E) whole-slide images without labels, producing embeddings that downstream models consume for classification, biomarker prediction, and survival analysis. The central motivation is that expert annotation of histopathology is scarce and expensive, while unlabeled slides are abundant; self-supervised pretraining on large, diverse tile corpora yields features that transfer strongly to clinical and research tasks with little labeled data.

The original Phikon, introduced in a medRxiv preprint in July 2023, was among the first works to systematically scale masked image modeling for histopathology. It demonstrated that an iBOT-pretrained Vision Transformer trained on pan-cancer tissue could match or exceed prior self-supervised approaches on weakly-supervised slide classification. Phikon-v2, released in September 2024, is the current generation: a larger ViT-L/16 trained with the DINOv2 recipe on an order of magnitude more data, positioned as a public feature extractor competitive with foundation models trained on proprietary slide collections.

Both models were released publicly on HuggingFace with accompanying code and substantive model cards, making Phikon one of the more openly documented pathology backbones despite a non-commercial license. Phikon-v2 is the recommended checkpoint for new projects, with the original Phikon retained as its predecessor.

#Key Features

• Two generations, one backbone family: Phikon (ViT-B, iBOT) and the current Phikon-v2 (ViT-L/16, DINOv2) share the same self-supervised, tile-level design, letting users choose a lighter or a larger, higher-performing extractor.
• Large public training corpus: Phikon-v2 was pretrained on PANCAN-XL, roughly 456 million H&E tiles drawn from about 60,000 whole-slide images sourced entirely from public cohorts including TCGA, CPTAC, GTEx, and TCIA, spanning more than 30 cancer sites.
• DINOv2 pretraining: Phikon-v2 combines DINO self-distillation with multi-crop, iBOT masked-image modeling, and KoLeo regularization, the modern recipe that has proven strong for dense pathology feature extraction.
• Biomarker-focused evaluation: Phikon-v2 was benchmarked on eight slide-level tasks with results reported on external validation cohorts, targeting molecular biomarker prediction from morphology alone.
• Openly released and reproducible: Model weights, inference code, and substantive HuggingFace model cards are public, though under Owkin's custom non-commercial license.

#Technical Details

Phikon-v2 is a Vision Transformer Large (ViT-L/16, patch size 16, embedding dimension 1024, approximately 0.3 billion parameters) pretrained with the DINOv2 self-supervised framework. Training ran for 100,000 iterations at batch size 4,096 across 32x4 NVIDIA V100 32GB GPUs (around 4,300 GPU hours) on the French Jean-Zay supercluster, using tiles of 224x224 pixels at 20x magnification. On its eight-task external benchmark suite, Phikon-v2 improved slide-level AUC by 1.75 points over single-shot retraining (p < 0.001), performing on par with proprietary histopathology foundation models and ahead of several of the 14 feature extractors it was compared against. The original Phikon is a ViT-B (about 85.8 million parameters) pretrained with iBOT on more than 40 million tiles from 16 TCGA cancer types, and reached state-of-the-art weakly-supervised slide classification across seven indications relative to ImageNet, MoCo v2, and other contemporary self-supervised methods.

#Applications

Phikon models function as frozen tile-level feature extractors within computational pathology pipelines. Researchers pass H&E tiles through the encoder and aggregate the embeddings with multiple-instance-learning heads for cancer subtyping, histological grading, and slide-level classification. A primary target of Phikon-v2 is molecular biomarker prediction, inferring markers such as microsatellite instability or mutation status from morphology alone to complement or triage molecular assays. The embeddings also support region-of-interest classification, tissue segmentation, and survival modeling, and serve as a public baseline backbone for benchmarking new pathology methods.

#Impact

Phikon helped establish self-supervised pretraining as the default approach for computational pathology backbones, and its open release lowered the barrier for groups without access to proprietary slide archives. Phikon-v2's demonstration that a model trained purely on public cohorts can rival proprietary foundation models is an important result for reproducibility in the field, and the family is widely used as a comparison baseline alongside UNI, Virchow, and CONCH. Key limitations are the H&E-centric training data, tile-level outputs that require a separate slide-level aggregation strategy, and the Owkin non-commercial license, which restricts commercial deployment. The models are research tools and have not received regulatory approval for clinical diagnostic use.