Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
The dataset viewer is not available for this split.
Server error while post-processing the rows. Please report the issue.
Error code:   RowsPostProcessingError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

CTFlow v2 — synthetic CT-RATE validation set

1,516 synthetic chest CT volumes, one per unique radiology report (Findings + Impressions, concatenated) in the CT-RATE validation split. Generated with the exact model and inference pipeline described in CTFlow v2: a connected latent space, and a week of very dumb bugs using the weights at huggingface.co/EnyaWoooo/ctflowv2-vlm3d2026.

Contents

  • *.nii.gz — 1,516 volumes, 512×512 in-plane, native frame count (typically 96–256), spacing 0.75×0.75mm in-plane with an adaptive z-spacing so every volume's block spans a fixed 384mm physical extent. HU clipped to [-1000, 1000]. Filenames match a representative VolumeName from the original CT-RATE validation_reports.csv that shares that report text (multiple original volumes can share one report; only one synthetic volume was generated per unique report).
  • prompts.json — the exact {"input_image_name", "report"} list used as input, so the report text behind every filename is reproducible without needing the original CT-RATE metadata.

How it was made

Each volume: 5 candidate autoregressive draws from independently verified-good seeds, jerk-median selection, and a CT-CLIP-based quality safety net (redraw with 2 backup seeds if the selected candidate scores below 0.5 against its own report text; publish whichever of the combined pool scores highest). Full pipeline details, including the real bugs found and fixed along the way, are in the write-up linked above.

Citation

See the model repo and github.com/WongJiayi/CTFlow-v2 for the citation entries.

Downloads last month
137