A Pan-Organ Vision-Language Model for Generalizable 3D CT Representations.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Foundation Models for Healthcare
• Radiology AI

Background:

• Existing AI models for computed tomographic (CT) imaging lack generalizability due to narrow training contexts.
• Limited anatomical coverage, contrast settings, and clinical indications restrict the real-world application of current AI tools.
• A need exists for AI models that can effectively process the wide spectrum of volumetric CT imaging data.

Purpose of the Study:

• To introduce Percival, a novel vision-language foundation model for CT medical imaging.
• To improve the generalizability of AI models in clinical radiology workflows.
• To assess the clinical relevance and latent structure discovered by Percival.

Main Methods:

• Developed Percival, a dual-encoder vision-language foundation model.
• Trained Percival on over 400,000 CT volumes and paired radiology reports from the Penn Medicine BioBank.
• Utilized a transformer-based image encoder and a BERT-style language encoder, aligned via symmetric contrastive learning.
• Validated Percival on over 20,000 participants' imaging data (100,000+ CT volumes).

Main Results:

• Percival demonstrated superior performance in image-text recall tasks compared to models trained on limited data.
• Evaluated Percival's clinical knowledge through association studies and survival analyses.
• Uncovered a rich latent structure within the data, aligning with physiological measurements and disease phenotypes.

Conclusions:

• Percival represents a significant advancement in generalizable AI for computed tomographic imaging.
• The model's ability to generalize across diverse CT data enhances its potential for improving clinical workflow efficiency.
• Percival's analysis revealed clinically relevant insights into biologic, phenotypic, and prognostic factors.