Benchmark of Intraoperative Ultrasound-to-MR Synthesis for Brain Tumor Surgery

Résumé fourni par la source

Trained model weights for the paper "A Systematic Benchmark of Intraoperative Ultrasound-to-MR Synthesis for Brain Tumour Surgery". This record contains the checkpoints that produced the published results. No intermediate training snapshots are included. Contents (66 files, 4.55 GiB). The gan/ directory (32 files, 1.87 GiB) holds Pix2Pix, SwinPix2Pix, CycleGAN and CUT: EMA generators for the 2D, 2.5D and full-3D regimes, plus the 3D refinement heads of the "2D + 3D-refine" variants, which reuse the matching 2D generator. The resvit/ directory (8 files, 1.14 GiB) holds the ResViT phase-2 best checkpoint (p2_best.pth) per experiment, which is the checkpoint used at inference. The syndiff/ directory (10 files, 0.31 GiB) holds the SynDiff diffusive generators at the evaluated epoch, plus the 3D refiners. The nnunet/ directory (16 files, 1.23 GiB) holds the two frozen downstream segmentation models (Seg-T2 and Seg-FLAIR), five folds each, with their plans and dataset descriptors. Every family covers both training targets: T2w-only and T2w + FLAIR multi-task. How to use them. The code is at https://github.com/smcch/ious2mr-benchmark. Clone the repository, then run "python scripts/fetch_weights.py --list" for the inventory, sizes and checksums, and "python scripts/fetch_weights.py --all" (or --family gan, resvit, syndiff or nnunet) to download. Then set IOUS2MR_CKPT to the weights directory. The file configs/weights_manifest.json in the repository maps each experiment name used in the paper to its file, size and SHA-256 checksum, and the download helper verifies them. Licence, please read before use. The gan/, resvit/ and nnunet/ directories are released under Apache-2.0. The syndiff/ directory is for non-commercial research use only: the SynDiff implementation derives from NVIDIA's DDGAN, distributed under the NVIDIA Source Code License, which restricts the work and any derivative works, including models trained with it, to research or evaluation purposes, and requires the same limitation to apply to anything you redistribute. Full attribution of every upstream component is in THIRD_PARTY_NOTICES.md in the repository. Provenance. Weights were trained on a single workstation (NVIDIA RTX 4080 SUPER 16 GB; a few ResViT runs on an RTX 3090), with the software versions pinned in envs/ in the repository. Re-training reproduces the reported trends but not bit-identical numbers, because the pipelines use non-deterministic GPU kernels. Data. Models were trained on ReMIND, The Cancer Imaging Archive (CC BY 4.0), https://doi.org/10.7937/3RAG-D070 Citation. If you use these weights, please cite the paper, the software (CITATION.cff in the repository) and the ReMIND dataset: Juvekar, P., Dorent, R., Kögl, F., et al. (2023). The Brain Resection Multimodal Imaging Database (ReMIND) (Version 1) [Data set]. The Cancer Imaging Archive. https://doi.org/10.7937/3RAG-D070

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.