LUS Consolidation Benchmark: model checkpoints
This collection contains 72 checkpoints: 36 segmentation models for each of the
512 x 512 and 224 x 224 tracks. Preserve the folder structure when downloading.
The downloaded directory is the --pth input for the code repository:
https://github.com/THA-NJU/LUS-Consolidation-Benchmark
weights_manifest.json maps every model/track to its tested checkpoint and SHA-256.
SHA256SUMS supports independent download verification. Checkpoints retain their
native serialization and must be loaded through the corresponding model adapter.
Use the code README's model-family environments; this is not one Transformers model.
Both tracks passed five-image evaluation, with five distinct test patients per track. This confirms the tested inference routes, not complete benchmark scores or clinical deployment validation. See the manuscript for benchmark results. The code's 35 representative minimal training checks do not establish full-run training reproduction or convergence.
The training dataset is not included. Dataset release is planned after paper acceptance. The 224 track was created by sliding crops and repartitioned by patient; the two tracks use independent patient splits.
Licensing is model-specific. The code repository's Apache-2.0 license must not be treated as a blanket license for all upstream models or derived checkpoints. Retain applicable upstream notices and terms, including the SAM License for SAM3. Consult the code repository's THIRD_PARTY_NOTICES.md and the relevant upstream source/weight licenses for each model before use or redistribution.