LUNA16-DiagnosticPreservation-2D (LUNA16-DP2D)

Dataset for Anatomical Attention-Based Medical Image Compression

---

Dataset Description

LUNA16-DP2D is a curated 2D axial CT slice dataset derived from the complete LUNA16 challenge (888 scans, subsets 0–9).

It is designed for research on diagnostic-preservation image compression — systems that allocate more bits to anatomically important regions (lungs, mediastinum, nodules) rather than optimizing purely for perceptual quality (PSNR/SSIM).

What's included

File pattern	Count	Description
slices/luna16_<subset>_<uid>_p<pct>.npy	6,216	Raw CT slices, int16 HU, 512×512
masks/lung_mask/<stem>_lung_mask.npy	6,216	Binary lung mask, uint8 0/1
masks/nodule_mask/<stem>_nodule_mask.npy	6,216	Binary nodule mask, uint8 0/1
metadata.csv	1	Per-slice metadata (see below)

Slice extraction

Slices extracted at percentiles [10, 25, 40, 50, 60, 75, 90] of each scan's axial depth (7 slices × 888 scans = 6,216 total):

idx = int(round(p / 100.0 * (depth - 1)))

Mask sources

Mask	Source	Labels
_lung_mask	seg-lungs-LUNA16 (binarized)	0=background, 1=lung+trachea
_nodule_mask	annotations.csv (ellipsoid projection)	0=background, 1=nodule

Mask verification

Check	Result
Binary values (0/1 only)	✅ all 6216
Lung coverage p25–p75	✅ 12–25% mean
Nodule hit rate	✅ 3.5% (220/6216 slices)
Nodules circular + inside lung	✅ confirmed visually

---

Usage

Load a slice + masks

import numpy as np
from pathlib import Path

DATASET = Path("/path/to/luna16-dp2d")

fname = "luna16_subset8_1.3.6.1...836860_p40.npy"
stem  = fname.replace(".npy", "")

ct          = np.load(DATASET / "slices" / fname)               # (512, 512) int16 HU
lung_mask   = np.load(DATASET / "masks" / f"{stem}_lung_mask.npy")   # (512, 512) uint8
nodule_mask = np.load(DATASET / "masks" / f"{stem}_nodule_mask.npy") # (512, 512) uint8

Load with PyTorch DataLoader

from medical_image_dataset_v4 import build_loaders

train_loader, val_loader = build_loaders(
    data_dir     = str(DATASET / "slices"),
    batch_size   = 8,
    modality     = "ct",
    patch_size   = 256,
    metadata_csv = str(DATASET / "metadata.csv"),
)

# Each batch contains:
# batch['image']      → [B, 3, 256, 256]  normalised [0,1]
# batch['filename']   → list[str]  original slice name
# batch['patch_row']  → [B]  patch row index
# batch['patch_col']  → [B]  patch col index

Build diagnostic importance map

from diagnostic_evaluator import DiagnosticEvaluator

evaluator = DiagnosticEvaluator(
    masks_dir  = str(DATASET / "masks"),
    patch_size = 256,
)

# In  training loop:

imp_map = evaluator.build_importance_map(
    ct_slices  = batch["image"],          # [B, 1, 256, 256]
    filenames  = batch["filename"],
    patch_rows = batch["patch_row"].tolist(),
    patch_cols = batch["patch_col"].tolist(),
)
# imp_map: [B, 1, 256, 256]
# 0.0=air | 0.3=soft tissue | 0.6=mediastinum | 0.8=lung | 1.0=nodule

CT normalization (HU windowing)

import numpy as np

def normalize_ct(arr):
    """Standard soft-tissue window: [-1000, 400] HU → [0, 1]"""
    arr = np.clip(arr.astype(np.float32), -1000.0, 400.0)
    return (arr + 1000.0) / 1400.0

---

metadata.csv columns

Column	Description
filename	Slice filename
subset	LUNA16 subset (subset0–subset9)
original_scan	Series UID
percentile	Percentile used for extraction (10/25/40/50/60/75/90)
slice_index	Exact axial index extracted
shape	Slice shape (always 512x512)
dtype	Array dtype
min / max / mean / std	HU statistics

---

License

CC BY 4.0 — derived from LUNA16 / LIDC-IDRI (CC BY 4.0).

You must cite both this dataset and the original LUNA16 paper (see below).

---

Citation

If you use this dataset, please cite:

@dataset{luna16_dp2d_2025,
  author    = {Nour El Houda, Bourai},
  title     = {{LUNA16-DiagnosticPreservation-2D (LUNA16-DP2D)}},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.XXXXXXX},
  url       = {https://doi.org/10.5281/zenodo.XXXXXXX},
  note      = {Derived from LUNA16 / LIDC-IDRI (CC BY 4.0)}
}

@article{setio2017luna16,
  title     = {Validation, comparison, and combination of algorithms for
               automatic detection of pulmonary nodules in computed tomography
               images: the LUNA16 challenge},
  author    = {Setio, Arnaud Arindra Adiyoso and others},
  journal   = {Medical Image Analysis},
  volume    = {42},
  pages     = {1--13},
  year      = {2017},
  publisher = {Elsevier},
  doi       = {10.1016/j.media.2017.06.015}
}

---

Related links

• 📦 Zenodo (citable, permanent): https://doi.org/10.5281/zenodo.XXXXXXX
• 🤗 HuggingFace (this page)
• 📊 Kaggle: https://www.kaggle.com/datasets/narita222/luna16-diagnostic-preservation-ct-dataset-2d
• 💻 Code: https://github.com/narita222/luna16-dp2d
• 🏆 LUNA16 Challenge: https://luna16.grand-challenge.org/