PRINT model weights
PRINT reconstructs 13 virtual protein representations (VPRs) from one or more grayscale ultrasound views of a thyroid nodule. This repository provides the weights for the first two training stages.
To produce the released Stage 1 checkpoint, we initialized the encoder with the official URFM weights (pretrained on multi-organ ultrasound images) and continued self-supervised training with a masked autoencoder on 4,666 ultrasound images from 1,616 thyroid nodules (follicular adenomas and follicular carcinomas) in discovery cohort B.
| Stage | Download | Contents and use |
|---|---|---|
| 1 ยท Adaptation to follicular thyroid neoplasms | Full MAE checkpoint | Epoch-150 masked autoencoder, including the encoder, mask token and pixel decoder. Use this checkpoint to initialize Stage 2. |
| 2 ยท VPR reconstruction | Ten fold checkpoints and manifest | Keep all ten safetensors files together with folds.json. The manifest records checkpoint hashes and the order of the 13 VPRs. |
For malignancy risk prediction, fit the Stage 3 diagnostic parameters separately using the PRINT training code.
Verify the downloads
Place the Stage 1 checkpoint, the contents of the Stage 2 directory and SHA256SUMS in one directory. From that directory run:
sha256sum -c SHA256SUMS
License and attribution
Both sets of model weights are licensed under the MIT License. The provenance notice identifies the upstream URFM checkpoint and describes its adaptation for PRINT. The original PRINT source code is licensed separately under Apache-2.0.
Declaration
PRINT is currently intended for research use only. Any clinical use must comply with the applicable laws and regulations of the country or region where the model is used.
Model tree for Jimmyzhe/PRINT
Base model
QingboKang/URFM