CONCH
A multimodal pathology foundation model trained using vision-language contrastive learning to align pathological images with pathology report text. The resulting image representations carry rich semantic information and can be used for zero-shot or transfer learning in downstream computational pathology tasks.
Timeline (last 90 days)
The study uses frozen pathology foundation model representations from UNI2-h and CONCH for feature extraction without fine-tuning.
Based on CONCH representations, the AP for colorectal cancer is 0.4410, outperforming DirectJoint's 0.3932; the AP for lung adenocarcinoma is 0.1659, higher than DirectJoint's 0.1226.
The frozen representations of pathology foundation models contain information for predicting rare joint molecular phenotypes, supporting such tasks without fine-tuning.
CONCH uses vision-language contrastive learning to align pathology images with pathology report text; UNI2-h is based on large-scale self-supervised pretraining on pathology images.
All Facts (4)
The study uses frozen pathology foundation model representations from UNI2-h and CONCH for feature extraction without fine-tuning.
50%UnverifiedBased on CONCH representations, the AP for colorectal cancer is 0.4410, outperforming DirectJoint's 0.3932; the AP for lung adenocarcinoma is 0.1659, higher than DirectJoint's 0.1226.
50%UnverifiedThe frozen representations of pathology foundation models contain information for predicting rare joint molecular phenotypes, supporting such tasks without fine-tuning.
50%UnverifiedCONCH uses vision-language contrastive learning to align pathology images with pathology report text; UNI2-h is based on large-scale self-supervised pretraining on pathology images.
50%