DeepLesion-1kTest3D

DeepLesion-1kTest3D Dataset Processing - Documentation

This document provides documentation for DeepLesion-1kTest3D - a 3D NIfTI conversion of 1,000 test cases from the DeepLesion dataset, enabling volumetric analysis of diverse lesion types across multiple anatomical regions.

📂 Dataset Information

Dataset Name: DeepLesion-1kTest3DSource: DeepLesion (NIH Clinical Center)Original Dataset Paper: Yan et al. (2018), "DeepLesion: automated mining of large-scale lesion annotations and universal lesion detection with deep learning"Original Format: 16-bit PNG slices (2D multi-slice representation)Converted Format: 3D NIfTI volumesTotal Test Cases: 1,000 3D CT volumesTotal Lesion Annotations: 4,927 lesions in test setProcessed By: Fakrul Islam TusharGitHub Repository: https://github.com/fitushar/HAIDZenodo NIfTI Repository: https://doi.org/10.5281/zenodo.18292965

Dataset Characteristics

Original DeepLesion: 32,735 lesions from 32,120 CT scans (10,594 patients)

Test Subset: 1,000 CT volumes with 4,927 lesion annotations

Lesion Types: 8 coarse categories (lung, liver, kidney, lymph node, bone, soft tissue, abdomen, pelvis)

Annotation Source: PACS system bookmarks from clinical radiology practice

Measurement: RECIST diameters (long/short axis measurements)

Slice Context: Variable slice range (typically 30-60 consecutive slices)

Test Set Organ-Specific Subsets

Liver Lesions: 701 liver lesion annotations (test set)

Kidney Lesions: 235 kidney lesion annotations (test set)

Other Organs: Lung, lymph nodes, bone, soft tissue, mediastinum

Coarse Lesion Type Distribution (Test Set)

Type 1: Lung nodules

Type 2: Abdomen lesions

Type 3: Mediastinum lesions

Type 4: Liver lesions

Type 5: Soft tissue lesions

Type 6: Kidney and urinary lesions

Type 7: Bone lesions

Type 8: Pelvic lesions

🔄 Processing Pipeline

1️⃣ PNG Slice Loading and 3D Volume Reconstruction

Purpose: Convert 2D 16-bit PNG slices to 3D NIfTI volumes with proper spacing

Input:

DeepLesion 16-bit PNG slices stored in patient folders

Format: Images_png/{Patient_Study_Series}/{SliceIndex}.png

Metadata: DL_info.csv with spacing and slice range information

Process:

import cv2
import numpy as np
import nibabel as nib

# Load consecutive slices
for slice_idx in slice_range:
im = cv2.imread(f'{slice_idx:03d}.png', -1) # -1 for 16-bit
# Remove intensity bias from 16-bit PNG encoding
im = (im.astype(np.int32) - 32768).astype(np.int16)
slices.append(im)

# Stack slices into 3D volume
volume = cv2.merge(slices) # or np.stack for >300 slices

# Create NIfTI with proper transformation matrix
spacing = [spacing_x, spacing_y, spacing_z] # mm per pixel
T = np.array([[0, -spacing[1], 0, 0],
[-spacing[0], 0, 0, 0],
[0, 0, -spacing[2], 0],
[0, 0, 0, 1]])
nifti_img = nib.Nifti1Image(volume, affine=T)
nib.save(nifti_img, output_path)

16-bit PNG Encoding:

DeepLesion stores CT in 16-bit PNG with intensity bias of +32768

Conversion subtracts 32768 to restore original Hounsfield Units (HU)

Resulting range: -1024 to +3071 HU (typical CT window)

Output:

3D NIfTI files: DeepLesion-1KTest-3D/{Patient_Study_Series}_SliceStart-SliceEnd.nii.gz

Format example: 000001_01_01_103-115.nii.gz (Patient 1, Study 1, Series 1, slices 103-115)

Total: 1,000 NIfTI volumes

2️⃣ Metadata and Lesion Annotation Integration

Purpose: Link 3D volumes with lesion annotations, measurements, and clinical metadata

Input:

Original DeepLesion metadata: DL_info.csv

Test split assignment

RECIST measurements (long/short axis)

Process:

Parse DeepLesion metadata CSV

Extract key slice index (slice with measurement annotation)

Map PNG file names to 3D NIfTI volumes

Preserve RECIST diameter measurements

Extract normalized lesion locations (0-1 coordinates)

Assign coarse lesion type labels (1-8)

Metadata Fields:

File_name: Original 2D PNG key slice (e.g., 000001_01_01_109.png)

Patient_index: Unique patient identifier (1-10594)

Study_index: Study number for patient (multiple studies per patient)

Series_ID: Series within study

Key_slice_index: Central slice with measurement annotation

Measurement_coordinates: RECIST long/short axis endpoints (x1,y1,x2,y2,x3,y3,x4,y4)

Bounding_boxes: Lesion bounding box (xmin,ymin,xmax,ymax)

Lesion_diameters_Pixel: Long and short axis lengths in pixels

Normalized_lesion_location: 3D normalized coordinates (0-1 range)

Coarse_lesion_type: Lesion category (1=lung, 2=abdomen, 3=mediastinum, 4=liver, 5=soft tissue, 6=kidney, 7=bone, 8=pelvis)

Possibly_noisy: Quality flag (0=clean, 1=potentially noisy annotation)

Slice_range: First and last slice indices in volume

Spacing_mm_px: Pixel spacing (x, y, z) in mm

Image_size: Image dimensions (typically 512×512)

DICOM_windows: Display window (level, width)

Patient_gender: M/F

Patient_age: Age at scan

Nifti_ct_name: Corresponding 3D NIfTI filename

Output Files:

DL_info_niftiID_test.csv (4,928 lesion annotations in test set)

test_Cases_list_DLplus_df.csv (801 unique test cases)

DeepLesion_Liver_test_set.csv (702 liver lesions)

DeepLesion_Kidneys_test_set.csv (236 kidney lesions)

3️⃣ Transformation Matrix for 3D Visualization

Purpose: Create proper affine transformation for correct display in 3D Slicer and ITK-SNAP

Transformation Matrix:

T = [[0, -spacing_y, 0, 0],
[-spacing_x, 0, 0, 0],
[0, 0, -spacing_z, 0],
[0, 0, 0, 1]]

Rationale:

DeepLesion uses axial slice ordering (superior to inferior)

Negative spacing ensures proper anatomical orientation

Compatible with 3D Slicer and ITK-SNAP viewers

Preserves patient coordinate system

Visualization:

3D Slicer: Load NIfTI volumes directly with correct orientation

ITK-SNAP: Lesion annotations can be overlaid using bounding boxes

Manual inspection: Verify lesion location matches key slice index

4️⃣ Organ-Specific Subset Extraction

Purpose: Create focused subsets for liver and kidney lesion research

Liver Lesion Subset:

Count: 701 liver lesion annotations (test set)

Coarse type: Label 4 (liver)

Use cases: Liver tumor detection, liver metastasis analysis

File: DeepLesion_Liver_test_set.csv

Kidney Lesion Subset:

Count: 235 kidney lesion annotations (test set)

Coarse type: Label 6 (kidney/urinary)

Use cases: Renal mass characterization, kidney cyst detection

File: DeepLesion_Kidneys_test_set.csv

Extraction Criteria:

# Liver lesions
liver_lesions = df[df['Coarse_lesion_type'] == 4]

# Kidney lesions
kidney_lesions = df[df['Coarse_lesion_type'] == 6]

📊 Dataset Statistics

Test Set Overview

Total NIfTI Volumes: 1,000 3D CT scans

Total Lesion Annotations: 4,927 lesions

Lesions per Volume: 1-10+ lesions (variable)

Unique Patients: ~400-500 patients (accounting for follow-up scans)

Slice and Spacing Characteristics

Slice Range: Typically 30-60 consecutive slices per volume

Shortest volume: ~12 slices

Longest volume: ~270 slices (extended abdomen/pelvis)

In-plane Spacing: 0.31 - 0.98 mm/pixel (variable resolution)

Slice Thickness: 1.0 - 5.0 mm (most common: 5mm)

Image Size: 512×512 pixels (consistent)

Lesion Size Distribution

Small lesions: <10mm diameter (~30%)

Medium lesions: 10-30mm diameter (~50%)

Large lesions: >30mm diameter (~20%)

RECIST measurements: Both long axis and short axis provided

Clinical Demographics

Age Range: 11-87 years (diverse adult and pediatric population)

Gender: Mixed male/female cohort

Clinical Context: Real-world radiology practice bookmarks

Anatomical Distribution (Test Set)

Liver (Type 4): 701 lesions (~14%)
Kidney (Type 6): 235 lesions (~5%)
Lung (Type 1): ~1,500 lesions (~30%)
Lymph Nodes (Type 3): ~800 lesions (~16%)
Bone (Type 7): ~600 lesions (~12%)
Soft Tissue (Type 5): ~500 lesions (~10%)
Abdomen (Type 2): ~400 lesions (~8%)
Pelvis (Type 8): ~191 lesions (~4%)

Image Quality Flags

Clean annotations: ~95% (Possibly_noisy=0)

Potentially noisy: ~5% (Possibly_noisy=1)

Noisy annotations: Ambiguous lesion boundaries or measurement inconsistencies

🗂️ File Structure

DeepLesion/
├── 036501_1.pdf # Original DeepLesion paper
├── DL_save_nifti.py # PNG-to-NIfTI conversion script
│
├── DL_info.csv # Full DeepLesion metadata (32,735 lesions)
├── DL_info_niftiID.csv # Metadata with NIfTI filenames
├── DL_info_niftiID_test.csv # Test set metadata (4,928 lesions)
│
├── test_Cases_list_DLplus_df.csv # Test set NIfTI volume list (801 cases)
├── val_Cases_list_DLplus_df.csv # Validation set NIfTI volume list
│
├── DeepLesion_Liver_test_set.csv # Liver lesion subset (702 lesions)
├── DeepLesion_Kidneys_test_set.csv # Kidney lesion subset (236 lesions)
│
├── DeepLesion-1KTest-3D/ # 3D NIfTI test volumes
│ ├── 000001_01_01_103-115.nii.gz # Patient 1, Study 1, Series 1
│ ├── 000001_03_01_058-118.nii.gz
│ ├── 000016_01_01_019-036.nii.gz
│ └── ... (1,000 NIfTI files)
│
└── DeepLesion-1KTest-3D.zip # Zenodo upload archive
└── Zenodo Repository: https://doi.org/10.5281/zenodo.18292965

🛠️ Software Dependencies

Core Libraries

opencv-python (cv2) 4.5.0+ (16-bit PNG reading)

nibabel 3.2.0+ (NIfTI creation and saving)

numpy 1.20.0+ (Array operations)

pandas 1.3.0+ (Metadata management)

Optional Visualization

3D Slicer 4.11+ (3D volume visualization)

ITK-SNAP 3.8+ (Lesion annotation overlay)

Installation

# Core dependencies
pip install opencv-python>=4.5.0 nibabel>=3.2.0 numpy>=1.20.0 pandas>=1.3.0

# Optional visualization tools
# Download 3D Slicer: https://download.slicer.org/
# Download ITK-SNAP: http://www.itksnap.org/

📈 Processing Highlights

PNG-to-NIfTI Conversion Advantages

3D Visualization: Enable volumetric viewing in 3D Slicer and ITK-SNAP

Spatial Context: Preserve anatomical relationships across slices

Standardized Format: NIfTI is widely supported by medical imaging tools

Metadata Preservation: Affine transformation encodes spacing and orientation

16-bit PNG Intensity Correction

Original: PNG stores values 0-65535

Bias: +32768 offset for signed 16-bit representation

Corrected: Subtract 32768 → -32768 to +32767 (Hounsfield Units)

Clinical relevance: Restored HU enables window/level visualization

Lesion Annotation Format

RECIST Measurements: Standard oncology response criteria

Bounding Boxes: Tight rectangular regions around lesions

Normalized Locations: 3D coordinates normalized to 0-1 range (body coverage)

Key Slice: Central slice with measurement annotation

Multi-Lesion Volumes

Many volumes contain multiple lesions (2-10+ lesions per scan)

Each lesion has independent annotation entry

Same NIfTI volume referenced multiple times in metadata

Enables comprehensive lesion detection benchmarking

🔗 Related Resources

Original DeepLesion Paper: Yan et al. (2018), IEEE J. Biomed. Health Inform.

DeepLesion Dataset: https://nihcc.app.box.com/v/DeepLesion

Zenodo NIfTI Repository: https://doi.org/10.5281/zenodo.18292965

Processing Code: https://github.com/fitushar/HAID

Processed By: Fakrul Islam Tushar

📝 Citation

If you use this pre-processed dataset, please cite:

Original DeepLesion Paper:

Yan, K., Wang, X., Lu, L., & Summers, R. M. (2018).
DeepLesion: automated mining of large-scale lesion annotations and universal lesion detection with deep learning.
Journal of Biomedical and Health Informatics, 22(4), 1091-1101.
DOI: 10.1109/JBHI.2017.2780066

DeepLesion Dataset:

Yan, K., et al. (2018).
DeepLesion: Large-scale Lesion Annotations from CT with Deep Learning.
NIH Clinical Center: https://nihcc.app.box.com/v/DeepLesion

DeepLesion-1kTest3D (3D NIfTI Conversion):

Fakrul Islam Tushar. (2026).
DeepLesion-1kTest3D: 3D NIfTI Conversion of DeepLesion Test Set.
Zenodo: https://doi.org/10.5281/zenodo.18292965
GitHub: https://github.com/fitushar/HAID

📌 Notes

3D Conversion: Complete PNG-to-NIfTI conversion for 1,000 test casesLesion Annotations: 4,927 diverse lesion annotations with RECIST measurementsOrgan-Specific Subsets: Liver (701) and kidney (235) lesion subsets availableTransformation Matrix: Optimized for 3D Slicer and ITK-SNAP visualization16-bit Intensity Correction: Proper Hounsfield Unit restoration (-32768 bias removal)Multi-Lesion Volumes: Many volumes contain multiple annotated lesionsZenodo Repository: All 1,000 3D NIfTI volumes publicly available at https://doi.org/10.5281/zenodo.18292965

Document Version: 1.0Last Updated: January 18, 2026Processed By: Fakrul Islam TusharContact: GitHub - https://github.com/fitushar/HAID