3DLAND: 3D Lesion Abdominal Anomaly Localization Dataset

This repository accompanies the 3DLAND project — a large-scale, organ-aware 3D lesion segmentation benchmark for abdominal CT scans — submitted for review to ACM Multimedia 2025.

🌐 Overview

• 6,000+ contrast-enhanced CT studies
• 3D lesion masks aligned with 7 abdominal organs
• Prompt-based annotation and propagation pipeline
• Applications: anomaly detection, lesion retrieval, organ-aware analysis

🧠 Pipeline

The lesion segmentation pipeline includes:

1. Organ segmentation via MONAI and lesion-to-organ assignment
2. 2D mask generation using SAM prompts
3. 3D mask propagation using MedSAM2

📦 Dataset

We curated over 6,000 contrast-enhanced abdominal CT scans from the publicly available DeepLesion dataset, selecting only those studies that include visible lesions or anomalies in abdominal organs.

To transform these raw scans into a structured, organ-aware 3D segmentation benchmark, we developed a multi-stage pipeline with both automated and expert-in-the-loop components:

1. Organ Segmentation: We used MONAI models trained on TotalSegmentator to segment seven abdominal organs — liver, kidneys, pancreas, spleen, stomach, and gallbladder. 1.2. Lesion-to-Organ Assignment: Lesions were matched to the most probable organ based on IoU overlap and 3D proximity, with ambiguous cases reviewed by clinicians.
2. 2D Lesion Mask Generation: Using MedSAM1, we generated lesion masks from DeepLesion’s bounding boxes. We found that shrinking the box to 70% of its original size, along with a center point prompt, significantly improved segmentation precision.
3. 3D Mask Propagation: The resulting 2D masks were propagated across slices using MedSAM2, producing dense 3D segmentations with anatomical continuity.

Each lesion in the dataset is:

• Annotated in 2D on the slice where the lesion is most clearly visible within the CT series
• Localized in 3D across all slices where the lesion is present and discernible
• Assigned to a specific abdominal organ
• Each 3D segmentation mask is saved as a stack of 2D PNG slices, preserving spatial consistency across the volume

The dataset includes:

• Phase II/2D_lesion_mask:2D lesion masks linked to organs
• Pahse III/3D_lesion_mask: 3D lesion masks linked to organs
• deeplesion_Info.csv: CSV file of Our dataset metadata according to DeepLesion metadata

All annotations underwent clinical review on 10–20% of lesions per organ to ensure high-quality ground truth.

📑 Metadata CSV Format

Each lesion in the 3DLAND dataset is described in a structured CSV file that includes key metadata for localization and segmentation. This CSV file links the image slices with their corresponding organ labels and bounding boxes.

📄 Sample Columns

Column	Description
series	Series ID from the DeepLesion dataset
slice_range	Range of axial slices in which the lesion may appear (according to deeplesion dataset metadata)
key_slice	Central slice with the most visible view of the lesion
lesion_id	Unique ID assigned to each lesion
matched_organs	Organ to which the lesion is anatomically linked
File_name	PNG file name of the key slice (e.g., 000002_02_01_050.png)
Bounding_boxes	Coordinates of the lesion in the key slice: [x_min, y_min, x_max, y_max]

---

📁 Folder Structure for Masks

📌 Phase II

• Location: Phase II/2D_lesion_mask
• Content: 2D lesion masks for the key slice
• File Naming: {series}_{key_slice}.png

📌 Phase III

• Location: Phase III/3D_lesion_mask
• Content: 3D volumetric lesion masks covering the entire slice range
• File Naming: {series}_K{key_slice}/{slice number}.png

---

⚠️ Notes on 3D Masks

• The 3D masks in Phase III cover the full slice range, but:
  • Only a few slices near the key slice contain non-zero segmentation.
  • Slices far from the key slice are fully black (all-zero).
• 💡 This design saves space while retaining anatomically relevant data since abdominal lesions are usually thin in the axial plane.

---

📌 Usage Tips

You can use the metadata CSV to:

• 🔍 Locate and crop lesions from the key slice for training or inference
• 🧠 Generate bounding box or center point prompts for segmentation models
• 🧬 Match each lesion to its organ for multi-organ anomaly analysis
• 🖼️ Load and visualize corresponding 2D or 3D masks

---

📂 Metadata File Location

• 3DLAND_Info.csv

---

📥 Downloading CT Scans

The original CT scan volumes used in 3DLAND are sourced from the DeepLesion dataset.

You can download the CT scans from the official NIH repository:

👉 Download CT scans from DeepLesion (NIH)

ℹ️ These CT volumes are required to visualize the lesion masks or to apply the 3DLAND segmentation pipeline. Make sure to match each series in the metadata with the corresponding CT study.

📄 License

The dataset and outputs are licensed under CC BY 4.0.
See the full license in the LICENSE file or at creativecommons.org/licenses/by/4.0