Phikon-v2, A large and public feature extractor for biomarker prediction: Open-source pathology foundation model with performance comparable to private models
In September 2024, Owkin and others released Phikon-v2, a pathology Vision Transformer (ViT-L) based on DINOv2 self-supervised learning, trained on over 460 million pathology patches from more than 100 public cohorts covering over 30 cancer types. On 8 slide-level tasks, Phikon-v2 outperforms its predecessor Phikon and achieves performance comparable to foundation models trained on private data, such as GigaPath and H-Optimus-0. The study also found that a simple ensemble strategy improves AUC by an average of 1.75 (p<0.001), and that small models can outperform large models on certain tasks (e.g., MSI prediction).
Phikon-v2 is one of the largest publicly available pathology foundation models to date, demonstrating that models trained entirely on public data can achieve performance comparable to those trained on private data. This lowers the barrier to entry for pathology AI research, enabling more institutions to develop diagnostic tools based on high-quality feature extractors. The study also reveals that model scale and downstream performance are not strictly positively correlated; small models may be superior on specific tasks, providing cost-effectiveness references for practical deployment.
Phikon-v2 uses the DINOv2 self-supervised learning framework to train a ViT-L architecture on 460 million pathology patches. Data comes from 100+ public cohorts covering 30+ cancer sites. Compared to Phikon (ViT-B), model size increases from 86M to 307M parameters. Evaluation covers 8 slide-level tasks, including cancer subtype classification and biomarker prediction (e.g., MSI, HRD), with all results reported on external validation sets to avoid data contamination. Comparing 14 feature extractors, Phikon-v2 achieves or approaches optimal performance on most tasks. Key findings: 1) DINOv2 is more suitable for scaling model and data size than iBOT; 2) simple ensembling (e.g., averaging multiple checkpoints) significantly improves performance; 3) the latest large models (GigaPath, H-Optimus-0) perform best overall but are surpassed by small models on tasks like MSI prediction.
The open-source release of Phikon-v2 will accelerate the application of pathology AI in drug development and clinical diagnosis. Pharmaceutical companies can use it to extract features and develop companion diagnostic kits. Hospital pathology departments can build localized AI-assisted diagnostic systems based on this model, reducing reliance on private data. Meanwhile, the study suggests that 'bigger is better' is not absolute; deployment should choose the appropriate scale based on the task.
Pathology AI companies can use Phikon-v2 as a feature extractor to develop classification or prediction models for specific cancer types, reducing pre-training costs. Cloud service providers can offer hosted inference APIs for Phikon-v2 to serve small and medium-sized hospitals and pharmaceutical companies. Investors can focus on pathology AI startups that leverage open-source foundation models to reduce R&D costs.
Future attention: 1) Validation results of Phikon-v2 in real clinical settings; 2) Whether the community develops fine-tuned versions based on this model; 3) Generalization ability in rare diseases or low-resource scenarios; 4) Cost-effectiveness comparison with larger models like GigaPath.