nielsRocholl/longi-registered

Longitudinal native CT (validation split)

Validation split (60 patients, 64 cases) of a longitudinal CT lesion dataset, in native (pre-registration, unwarped) frame. Baseline (BL) and follow-up (FU) scans are each in their own original acquisition space — no image registration has been applied.

This replaces a prior push that shipped uniGradICON-registered/warped-frame data with incomplete per-timepoint click files. This native version is the corrected eval package: every case now has full baseline and follow-up scans, masks, and click prompts with correct union-of-lesions semantics (see below).

Contents

• inputsTrBL/ — baseline CT (.nii.gz) + click-prompt JSON (.json), native BL frame
• inputsTrFU/ — follow-up CT (.nii.gz) + click-prompt JSON (.json), native FU frame
• targetsTrBL/ — baseline lesion segmentation mask, native BL frame
• targetsTrFU/ — follow-up lesion segmentation mask, native FU frame
• meta/<patient_id>.csv — per-patient lesion metadata: lesion_id, cog_bl, cog_backpropagated, img_id_bl, cog_propagated, cog_fu, img_id_fu, lesion_type, topology_class, merged_into, volume_bl, volume_fu, target_lesion, use_for_challenge, linking_unclear
• val_patients.csv — the 60 patient IDs defining this split

Case naming

<patient_id>_<scan_idx> — scan_idx matches img_id_bl / img_id_fu in the corresponding patient's meta/<patient_id>.csv.

Click JSON semantics

Each click JSON's point list is the full union of lesions present at either timepoint for that case, filtered to use_for_challenge == True. Point "name" equals lesion_id, paired 1:1 across the BL and FU json for a case (same lesion_id appears at the matching point index in both).

• inputsTrBL/<case>.json: point = cog_bl if present, else cog_backpropagated (a back-propagated stand-in location for lesions that only appear in follow-up, i.e. topology_class == NEWLYAPPEARING, so they still get a plausible BL-frame click).
• inputsTrFU/<case>.json: point = cog_propagated (baseline location propagated forward) if present, else native cog_fu (covers lesions that disappear by follow-up, i.e. topology_class == DISAPPEARING, giving them an FU-frame click even though there's no visible lesion there anymore).

A lesion is only included in a given case's BL (or FU) json if its img_id_bl (or img_id_fu) matches that case's scan_idx — a lesion can be tracked across different scan indices at BL vs FU for the same patient, in which case it belongs to different case files.

Citation

Derived from the Universal Lesion Segmentation longitudinal extension. See project repo for dataset construction methodology.