ResEncL-HighIQS-MAE

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

Model ID: 752; BaseMAETrainer; final checkpoint at epoch 1000. No downstream fine-tuned weights are included. This is a native PyTorch encoder, not a Transformers AutoModel.

Status and rights

The encoder weights are released under Creative Commons Attribution 4.0 International, with team permission confirmed by the author. Please credit the paper and this model repository and indicate modifications. Architecture reference: MIC-DKFZ/nnssl. The loader is adapted from its ResEncL architecture configuration; the local source snapshot carries CC-BY-SA-4.0, whose attribution and share-alike obligations are retained separately for that code. The weight license does not relicense third-party dependencies.

Input and usage

Single-channel 3D input, with the recorded pretraining plan using 1 mm isotropic resampling, Z-score normalization and recommended 160-cube patches. Input preparation is external to this loader. The exact training subset definition and preprocessing conventions should be consulted before downstream adaptation.

from load_encoder import load_encoder
encoder = load_encoder('.')
# Provide your preprocessed tensor [batch, 1, depth, height, width].

For a synthetic 64-cube loading check: python load_encoder.py .. This does not establish clinical validity or downstream performance. See validation.json for strict loading, exact roundtrip and finite-value checks.

Only encoder tensors and necessary configuration are present. Reconstruction decoder, optimizer, training logs, initialization arguments and training sample manifests are excluded. Intended for research; not validated as a clinical diagnostic system. This is an intended-use statement, not an additional restriction on the CC BY license. No later Brier/selection development results are claimed as v1 results.