radiology-annotation

ROI / VOI / Mask Annotation SOP

Use this skill to make the annotation behind a radiomics/segmentation/radiogenomics study defensible. Reviewers reject papers when the segmentation is a black box: unknown readers, no reproducibility, masks that don't align with the images, or feature instability never tested. This skill specifies the lesion-selection strategy, the reader protocol, the reproducibility plan, and the geometric checks — and writes the Methods text.

Core stance

• Segmentation is a measurement, so characterise its error. Report inter-/intra-observer reproducibility and propagate it (drop unstable features) — don't treat masks as ground truth.
• Pre-specify the lesion-selection rule. 2D vs 3D, whole-tumour vs largest-slice vs peritumoral ring vs sub-regional habitat vs multi-lesion handling — decided up front, applied uniformly, and justified by the biology and the endpoint.
• Readers are part of the method. Number, seniority, blinding to outcome, independent vs consensus, and the adjudication rule for disagreement all belong in Methods.
• Geometry must be exact. Mask and image must share spacing, origin, direction, and slice order; a one-voxel or flipped-axis mismatch silently corrupts every feature.
• Reproducibility before modelling. Filter to reproducible features (e.g. ICC threshold) before selection/modelling, on training data only — leakage hides here too.
• Integrity. Never invent ICC/Dice values, reader counts, or QC results; mark what must be measured.

When to use

• "Design an ROI/mask annotation SOP / 帮我写标注 SOP（勾画规范）。"
• "Two radiologists contoured the lesions — how do I report agreement and select stable features?"
• "Whole-tumour or largest-slice? 2D or 3D? peritumoral? habitat?"
• "My masks don't line up with the images / DICOM-SEG/RTSTRUCT conversion issues."
• "Write the segmentation + reproducibility paragraph for Methods."

When to open extra files