Confidence-gated patch selection for volumetric photoacoustic fusion.
This is a restricted, pre-acceptance source preview for inspection only. It is not an open-source release. Running the code, training models, performing inference, or using these source files to reproduce experiments is not authorized under the current release terms. See LICENSE.txt.
An implementation release under a separately announced license is planned after paper acceptance. Pretrained weights will not be distributed. Public visibility does not grant permission to execute or reproduce this preview.
The selector extracts compact numerical summaries from paired three-dimensional patches. It submits source-exchange-invariant descriptors to a typed decision interface, combines utility and risk estimates, and applies a confidence gate. Low-confidence decisions use a deterministic fallback. Selected patch coordinates are cached before fusion-network training; fusion inference does not call the decision service.
| Directory or file | Purpose |
|---|---|
selection/jev_block_selection.py |
Patch descriptors, typed requests, deterministic ranking, and confidence fallback. |
selection/prepare_jev_requests.py |
Request preparation from a caller-supplied manifest. |
selection/run_openjev_requests.py |
Response processing and gated/ungated coordinate caches. |
fusion/data.py |
Volume loading, split consistency checks, and patch dataset. |
fusion/model.py |
Three-dimensional backbone, attention, reliability gating, and symmetric fusion. |
fusion/losses.py |
Explicit supervision and optional research losses. |
fusion/train.py |
Configuration-driven training, validation, and checkpoint handling. |
fusion/infer.py |
Weight-based fusion inference. |
The source uses NumPy, SciPy, and PyTorch. They are separate third-party dependencies and are not bundled. OpenJev is also an independent upstream dependency, not included in this archive.
This archive contains no pretrained or trained weights, private acquisition data, teacher volumes, labels, real sample manifests, actual experiment configurations, decision responses, selection caches, or training logs. It contains no installation or reproduction instructions.
No author identity, affiliation, personal contact information, personal filesystem paths, remote server addresses, or credentials are included. The preview has not been packaged from the original Git history.
Omitting experimental artifacts limits direct reproduction of the reported experiments. It does not provide a technical guarantee against independent implementation or retraining. The restriction on using this preview is a release condition, not a technical execution barrier.