3DLAND: 3D Lesion Abdominal Anomaly Localization Dataset

TL;DR

3DLAND addresses the lack of large-scale multi-organ 3D lesion datasets by releasing a benchmark with over 6,000 contrast-enhanced abdominal CT volumes and 20,000+ organ-linked 3D lesion annotations across seven organs. The authors introduce a three-phase pipeline that combines automated spatial reasoning for lesion-to-organ assignment, prompt-optimized 2D segmentation, and memory-guided 3D propagation to reconstruct voxel-accurate organ-aware lesions, validated by expert radiologists. The work establishes a robust benchmark for anomaly detection, localization, and cross-organ transfer learning, and demonstrates strong performance and reliability relative to prior resources. The dataset and accompanying code are publicly available under CC BY 4.0 to enable reproducibility and foster further research in representation learning and clinical AI applications.

Abstract

Existing medical imaging datasets for abdominal CT often lack three-dimensional annotations, multi-organ coverage, or precise lesion-to-organ associations, hindering robust representation learning and clinical applications. To address this gap, we introduce 3DLAND, a large-scale benchmark dataset comprising over 6,000 contrast-enhanced CT volumes with over 20,000 high-fidelity 3D lesion annotations linked to seven abdominal organs: liver, kidneys, pancreas, spleen, stomach, and gallbladder. Our streamlined three-phase pipeline integrates automated spatial reasoning, prompt-optimized 2D segmentation, and memory-guided 3D propagation, validated by expert radiologists with surface dice scores exceeding 0.75. By providing diverse lesion types and patient demographics, 3DLAND enables scalable evaluation of anomaly detection, localization, and cross-organ transfer learning for medical AI. Our dataset establishes a new benchmark for evaluating organ-aware 3D segmentation models, paving the way for advancements in healthcare-oriented AI. To facilitate reproducibility and further research, the 3DLAND dataset and implementation code are publicly available at https://mehrn79.github.io/3DLAND.

Figures (6)

• Figure 1: Comprehensive overview of 3DLAND dataset properties, highlighting its superiority in anomaly lesion coverage and demographic diversity.Property I: Number of anomaly lesions per abdominal organ, visualized with a 3D anatomical model showing lesion counts for liver, gallbladder, pancreas, spleen, kidneys, and stomach.Property II: Significant demographic diversity, including gender distribution and age groups, alongside key statistics.Property III: Organ-based comparison of 3DLAND anomaly 3D lesion counts with the largest datasets per each organ, demonstrating 3DLAND's extensive coverage across organs. This figure underscores 3DLAND's value as a large-scale, organ-aware benchmark for abdominal CT anomaly localization and downstream ML tasks.
• Figure 2: Overview of our three-stage pipeline for generating organ-aware 3D lesion masks from DeepLesion CT data. Phase I assigns each lesion to its most likely abdominal organ using TotalSegmentator wasserthal2023totalsegmentator organ segmentation and spatial overlap reasoning. Phase II performs precise 2D segmentation using a prompt-optimized MedSAM1 ma2024segment model. Phase III leverages MedSAM2 ma2025medsam2 to propagate key-slice masks across volume slices, resulting in anatomically consistent 3D lesion reconstructions.
• Figure 3: Effect of bounding box shrinkage on 2D lesion segmentation with MedSAM1. As the box size decreases from 100% to 70%, the pseudo mask (red) aligns more closely with the ground truth (blue), and the error region (yellow) is reduced. The 70% box provides the best trade-off between focus and coverage.
• Figure 4: Performance of MedSAM1 across bounding box scales (100–60%) on three metrics: Dice, IoU, and Surface Dice. All metrics peak at 70% scale, confirming that moderately shrinking prompts improves lesion boundary accuracy by reducing background noise.
• Figure 5: Evaluation of MedSAM Model for Lesion Mask Prediction in Multi-Organ Abdominal CT Scans. This figure presents a comparative analysis of lesion mask predictions generated by the MedSAM on four real-world clinical cases from abdominal CT scans. Each case row (Cases 1--4) displays four axial slices with overlaid annotations to evaluate prompting strategies for semi-automatic segmentation. The green contour highlights the target organ, while prediction masks are color-coded as follows: Yellow Bounding box (bbox) only, providing coarse spatial guidance; Blue: Bounding box + center point (positive prompt inside the lesion), refining localization; Red bounding box + negative points (prompts outside the lesion to exclude false positives); Orange: Center point + negative points (point-based prompting with exclusion cues). Key Observations: Improved Accuracy with Center Point Prompting---Across all cases, the bbox + center point (blue) strategy yields the most precise masks, closely aligning with lesion boundaries (e.g., in Case 1's liver tumor and Case 3's renal mass). This hybrid approach leverages the bbox for broad context and the center point for focal refinement, reducing over-segmentation seen in bbox-only (yellow) predictions. Adverse Effect of Negative Points---Incorporating negative points (red and orange) degrades performance, leading to under-segmentation or fragmented masks (e.g., incomplete coverage in Case 2's lesion and erratic boundaries in Case 4's hepatic abnormality). This suggests negative prompts introduce ambiguity in MedSAM's attention mechanism, particularly in heterogeneous tissues. Clinical Implications: These samples demonstrate MedSAM's potential for efficient, interactive segmentation in radiology workflows, but prompt design is critical---favoring positive, anatomically informed cues over exclusionary ones to enhance diagnostic reliability.