AbdomenAtlas: A large-scale, detailed-annotated, & multi-center dataset for efficient transfer learning and open algorithmic benchmarking

2024-07-01

Medical Image Anal.

AI Summary

Existing abdominal CT segmentation datasets suffer from limited scale, coarse-grained annotations, and single-center bias, hindering transfer learning and fair algorithm benchmarking. To address this, we introduce AbdomenAtlas—the first large-scale, multi-center abdominal CT dataset featuring dual-level (organ and lesion) fine-grained 3D pixel-wise annotations across 7 hospitals and 5 scanning protocols, covering 33 anatomical structures and pathologies. We propose a semi-automatic annotation paradigm integrating weakly supervised pre-labeling, expert-guided 3D interactive refinement, standardized DICOM preprocessing, and privacy-preserving de-identification. Evaluated on eight downstream tasks, models trained on AbdomenAtlas achieve an average 4.2% mDice improvement and 37% higher few-shot transfer stability. The dataset is publicly released and adopted as the official MICCAI 2024 benchmark.

Technology Category

• Computer Vision: Segmentation
• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning
• Knowledge Representation and Reasoning: Diagnosis and Abductive Reasoning

Application Category

• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets
• Security and Privacy: Data transparency and provenance
• Semantics and Knowledge: Scalable techniques for the creation, curation, publication, maintenance, and consumption of large, Web-based, structured, reusable, knowledge graphs and ontologies

Problem

• Creating a large-scale abdominal CT dataset for AI training
• Reducing radiologists' workload via semi-automatic annotation
• Establishing benchmark for AI algorithm evaluation

Innovation

• Large-scale multi-center abdominal CT dataset
• Semi-automatic AI-assisted annotation procedure
• Open benchmark for AI algorithm evaluation