Detailed 3D Scans of over 6,000 Patients Boost Accuracy in Detecting Abdominal Lesions

Researchers are tackling the limited availability of comprehensively annotated three-dimensional medical imaging datasets for abdominal computed tomography. Mehran Advand, Zahra Dehghanian, and Navid Faraji, all from Sharif University of Technology, alongside colleagues Reza Barati, Seyed Amir Ahmad Safavi-Naini, and Hamid R. Rabiee, have introduced 3DLAND, a new large-scale benchmark dataset designed to overcome these challenges. This dataset comprises over 6,000 contrast-enhanced CT volumes and features more than 20,000 high-fidelity three-dimensional lesion annotations across seven key abdominal organs, liver, kidneys, pancreas, spleen, stomach, and gallbladder. The significance of this work lies in its potential to enable scalable evaluation of anomaly detection, localisation, and cross-organ transfer learning, ultimately advancing the development of healthcare-oriented artificial intelligence and establishing a new standard for organ-aware 3D segmentation models.

Comprising over 6,000 contrast-enhanced CT volumes and more than 20,000 meticulously annotated 3D lesions, 3DLAND addresses a critical limitation in the field, the lack of comprehensive, multi-organ datasets with precise lesion details.

The creation of 3DLAND involved a novel three-phase pipeline integrating automated spatial reasoning with advanced 2D segmentation techniques and memory-guided 3D propagation. This streamlined process ensures high-fidelity annotations, validated by expert radiologists who confirmed the accuracy of the 3D lesion masks with surface dice scores exceeding 0.75.

By providing detailed volumetric data and linking lesions to specific organs, 3DLAND facilitates the development of more robust and clinically relevant AI models0.3DLAND’s diverse lesion types and patient demographics further enhance its value, enabling scalable evaluation of anomaly detection, localisation, and cross-organ transfer learning. Current benchmarks often focus on single organs or lack detailed lesion information, hindering the development of algorithms capable of generalising across the entire abdominal region.

Researchers anticipate that 3DLAND will serve as a new benchmark for evaluating organ-aware 3D segmentation models, ultimately paving the way for advancements in healthcare-oriented AI and improved patient outcomes. The dataset and associated implementation code are publicly available, fostering reproducibility and encouraging further research within the medical imaging community.

The 3DLAND dataset facilitates the construction of robust, organ-aware 3D lesion annotations, achieving an average Surface dice score of 0.755 on validated 3D masks. This performance was measured on a 10% subset comprising 600 CT volumes, assessed against manual annotations performed by a medical expert. Processing over 6,000 CT volumes required approximately 10 hours of computation using a single NVIDIA A100 GPU, equating to 0.05 GPU hours per volume and demonstrating scalability for large-scale applications.

Phase I, focused on lesion-to-organ assignment, achieved 94.8% assignment accuracy on a 10% subset of 300 volumes, as validated by a medical expert utilising spatial reasoning with IoU exceeding 10% and Euclidean distance below 20 pixels. Ablation studies demonstrated that these thresholds minimise ambiguous assignments, limiting them to 15% while maintaining high accuracy.

Assignment Accuracy specifically measures the percentage of lesions correctly linked to one of seven abdominal organs, gallbladder being critical for both clinical diagnosis and machine learning tasks. TotalSegmentator, employed as the base model, achieved Dice scores ranging from 82.0 to 94.5% across all organs, with low variance between 1.2 and 1.8.

Notably, the model excelled on larger organs like the liver (94.5 ±1.2) and kidneys (94.0 ±1.3), while also delivering strong results on smaller organs such as the pancreas (82.8 ±1.8) and gallbladder (92.5 ±1.7). Compared to alternative state-of-the-art models, including MedSAM2, TotalSegmentator consistently outperformed them by 3, 5%, particularly in complex cases like the gallbladder where it achieved a score of 85.0 ±2.8.

A three-phase pipeline underpinned the creation of 3DLAND. Initially, automated spatial reasoning was employed to assign each detected lesion to its most probable abdominal organ, utilising the TotalSegmentator for precise organ segmentation. This approach leverages the anatomical proximity of lesions to organs, establishing a crucial link between anomalies and their location within the abdomen.

Subsequently, a prompt-optimised MedSAM model was implemented for accurate 2D segmentation of lesions on individual slices, leveraging its capabilities in prompt-based segmentation to refine delineation of lesion boundaries with minimal manual intervention. Following 2D segmentation, a memory-guided 3D propagation technique was applied to extend key-slice masks across the entire CT volume, intelligently interpolating between slices to reconstruct a complete 3D representation of each lesion while preserving anatomical consistency.

To ensure clinical reliability, expert radiologists validated the generated 3D annotations, achieving Dice scores exceeding 0.75. The persistent challenge of reliably detecting subtle disease in abdominal scans has long frustrated the potential of artificial intelligence in healthcare. For years, algorithms have struggled not simply with seeing anomalies, but with understanding where those anomalies lie within the complex anatomy of the abdomen and, crucially, linking them to the correct organ.

Existing datasets, while plentiful, often provide only two-dimensional labels or lack the comprehensive, multi-organ coverage needed to train truly robust systems. The release of 3DLAND represents a significant step towards addressing this bottleneck, focusing on precise, three-dimensional lesion mapping coupled with clear organ associations, allowing for a new level of algorithmic training and evaluation.

Researchers can now build models that don’t just identify something is wrong, but where it is wrong, within a specific anatomical context, vital for applications ranging from automated diagnosis to surgical planning and even robotic-assisted interventions. However, the reliance on contrast-enhanced CT scans introduces a limitation, as the dataset may not generalise well to imaging modalities like MRI or ultrasound.

Furthermore, while the validation scores are promising, real-world clinical variability is notoriously difficult to capture fully in any dataset. The observed sensitivity of the MedSAM model to prompt design highlights the need for continued investigation into how best to interface AI with radiologists’ expertise.