Reference segmentations of tumour and resection cavity for the ReMIND cohort (ioUS-to-MR synthesis benchmark)

Résumé fourni par la source

Manual reference segmentations of tumour and resection cavity for the ReMIND cohort, created for the downstream evaluation of the paper "A Systematic Benchmark of Intraoperative Ultrasound-to-MR Synthesis for Brain Tumour Surgery". They cover the pre-resection and post-resection studies of the cohort. Contents: 297 label maps from 77 subjects. The mr_space/ directory holds 150 files, drawn on the pre-processed MR (T2w / FLAIR) after co-registration to the intraoperative ultrasound reference and cropping to the ultrasound field of view. The us_space/ directory holds 147 files, the corresponding labels on the intraoperative ultrasound grid. Files are named ReMIND- - -{mri,us}-segmentation.nii.gz, where is "pre" (pre-resection) or "post" (post-resection). Labels: 0 is background, 1 is tumour, 2 is resection cavity. In the paper the two foreground classes are also analysed together as a single lesion class (tumour plus cavity), which is the primary downstream endpoint. How they were produced. Labels were generated semi-automatically in ITK-SNAP with the nnInteractive plug-in, then reviewed and refined by a board-certified neurosurgeon. They were drawn directly on the pre-processed volumes, so that real and synthetic MRI inputs share an identical spatial reference. An inter-observer subset, annotated by two independent observers, is analysed in the paper to establish a human agreement ceiling; the scripts are src/downstream/interobserver_MR.py and src/downstream/interobserver_US.py in the repository. Intended use. These labels let others reproduce the downstream segmentation evaluation, train their own tumour and cavity models on ReMIND, or benchmark intraoperative ultrasound segmentation without repeating the annotation effort. To reproduce the protocol of the paper exactly, use configs/subject_split.json in the repository: 61 training and 16 held-out subjects, split at the subject level. Code: https://github.com/smcch/ious2mr-benchmark Source dataset: ReMIND, The Cancer Imaging Archive, https://doi.org/10.7937/3RAG-D070 Licence. ReMIND is distributed under CC BY 4.0, which permits redistribution of derived works with attribution. These segmentations are a derived work and are released under the same licence. Required attribution: 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 Please also cite the paper these annotations were made for.

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