Model Card for Pillar0-HeadCT

Model Details

Model Description

Pillar-0 (Head CT) is a general-purpose radiology foundation model designed for volumetric Head CT analysis. It leverages the Pillar-0 architecture (Atlas Vision Encoder + Qwen3-Embedding-8B Text Encoder) to process full 3D volumes, enabling the detection of critical neurological findings that require spatial context.

This model is particularly notable for its data efficiency. In downstream tasks like intracranial hemorrhage detection (RSNA), Pillar-0 achieved >95 AUROC using only 1/20th of the training data required by the next most efficient baseline.

• Model type: Vision-Language Foundation Model.
• Architecture: Atlas Vision Encoder (Backbone) aligned with Qwen3-Embedding-8B (Text Encoder).
• Language(s) (NLP): English (Radiology Reports).

Model Sources

• Repository: https://huggingface.co/collections/YalaLab/pillar-0
• Code: https://github.com/YalaLab/pillar-pretrain
• Paper: Pillar-0: A New Frontier for Radiology Foundation Models

Uses

• Acute Pathology Detection: Highly effective for detecting intracranial hemorrhage (ICH), mass effect, hydrocephalus, and herniation.
• Data-Efficient Fine-tuning: Ideal for training specialist models in scenarios with limited labeled data (e.g., rare neurological conditions).

Evaluation

Testing Data & Metrics

• Test Set: 4,906 exams (UCSF).
• Protocol: Linear probing via RATE-Evals on 29 clinically grounded findings (e.g., hemorrhage, fractures, post-op changes).

Results

Pillar-0 achieves the highest performance gap over baselines in the Head CT modality.

Model	Mean AUROC	Win Rate vs Pillar-0
Pillar-0 (Ours)	90.1	-
MedGemma	80.2	0.0%
MedImageInsight	78.3	6.9%
Merlin	69.9	0.0%

Citation

@article{pillar0,
  title   = {Pillar-0: A New Frontier for Radiology Foundation Models},
  author  = {Agrawal, Kumar Krishna and Liu, Longchao and Lian, Long and Nercessian, Michael and Harguindeguy, Natalia and Wu, Yufu and Mikhael, Peter and Lin, Gigin and Sequist, Lecia V. and Fintelmann, Florian and Darrell, Trevor and Bai, Yutong and Chung, Maggie and Yala, Adam},
  year    = {2025}
}