DINO3D-ResEncL-NoGram-E260

Encoder-only release of model 772 accompanying Benchmarking Self-Supervised 3D Medical Vision Foundation Models and Transferability Estimation.

Which weights are released?

This is the epoch-260 student encoder in network_weights from openmind_resencl_nogram_epoch260.pth. This is the state selected by the benchmark's load_cnn_dino_encoder loader. The EMA teacher_encoder and both projection heads are excluded. Student and teacher must not be interchanged when reproducing model 772.

This is a custom native-3D DINO-style CNN with a ResEncL-style six-stage layout; it is not the standard nnssl ResEncL architecture and is not an official Meta DINOv3 ViT checkpoint. The recorded configuration has no initialization checkpoint. The training code's generic variant field does not change the actual CNN architecture.

Input and outputs

Single-channel [B,1,D,H,W] input. The recorded pretraining view is nnsslPlans_onemmiso, with 160-cube global patches and 96-cube local crops. The model772 benchmark plan uses 1 mm isotropic resampling and Z-score normalization. Preprocessing is external to the loader; consult the paper's data/view definitions before adapting to another modality or task.

The six native feature maps have channels [32,64,128,256,320,320] and cumulative strides [1,2,4,8,16,32]. Blocks per stage are [1,3,4,6,6,6]. This release contains 90,398,529 parameters, including the native mask token. No downstream segmentation decoder is provided.

Usage

Download the repository files, install requirements.txt, and run:

from load_encoder import load_encoder
encoder = load_encoder('.')
# outputs = encoder(preprocessed_tensor)  # six 3D feature maps

python load_encoder.py . runs a synthetic 64-cube loading check. This is a native PyTorch model, not a Transformers AutoModel. The loader uses safetensors and strict state loading.

Validation and provenance

validation.json records the original checkpoint SHA256, exported encoder SHA256, exact tensor roundtrip, finite weights and finite synthetic forward outputs. The bundled dino_cnn_encoder.py matches the original team source in dinov3-3d-nnunet/src/dino3d/cnn_encoder.py and the classification handoff copy. Only encoder tensors and necessary public configuration are included; optimizer, training sample lists, logs, teacher, projection heads and downstream weights are excluded.

License and limitations

The team-authorized encoder weights and accompanying team-authored architecture/loader code are released under CC BY 4.0; credit the paper and this repository and indicate modifications. Third-party dependencies retain their own licenses. Intended for research; no clinical diagnostic validation is claimed. This is an intended-use statement, not an additional restriction on the license. No later transferability-development results are claimed as arXiv v1 results.