---

name: data-preparation description: "Prepare and validate ISIC, BRATS, and custom 2D/3D image-segmentation data for the MedSegDiff dataset loaders." disable-model-invocation: true metadata: disco-role: operating license: MIT

Data preparation

Use this sub-skill when a MedSegDiff run needs an input directory, a loader choice, or a diagnosis of missing/misaligned image and mask data. Data is not bundled with this skill; acquire it separately and keep the split directory layout described in data-formats.md.

Route by dataset

• ISIC: use ISICDataset; first resolve the source-code CSV-name versus README CSV-name mismatch. See data-formats.md.
• BRATS: use BRATSDataset for full-volume tensors or BRATSDataset3D for the repository's fixed 155 virtual slices per case.
• Custom 2D: use CustomDataset with images/*.png and masks/*.png.
• Custom 3D: use CustomDataset3D with paired NIfTI volumes and account for its constructor and launcher integration caveat.

Read api-reference.md before writing a caller or transform. Read troubleshooting.md when a constructor assertion, empty dataset, shape error, or unexpected tensor value appears. The bundled validator is a conservative layout check, not a replacement for loading every dataset-specific file:

python scripts/validate_dataset_layout.py DATA_ROOT --kind custom2d
python scripts/validate_dataset_layout.py SPLIT_ROOT --kind isic --mode Training
python scripts/validate_dataset_layout.py CASE_ROOT --kind brats --mode 3d

Preparation invariant

Before handing data to training or sampling, identify the exact loader and confirm its tuple contract, modality set, spatial crop/resize policy, and mask binarization policy. Do not silently repair filenames, pair lists by basename, or normalize medical volumes: those are behavior changes that require an explicit loader extension.

This sub-skill covers data discovery and loader behavior only. Route optimizer, checkpoint, training flags, sampling, and evaluation to the sibling skills.