TL;DR
This paper introduces the DAP Atlas, a novel, comprehensive dataset of full-body CT scans with 142 anatomical labels, generated through an automated knowledge aggregation process. This dataset and its associated robust segmentation model aim to overcome the limitations of existing, fragmented medical imaging datasets, enabling more holistic anatomical analysis and downstream applications.
What Problem Does This Paper Address?
Medical image segmentation has seen significant advancements, largely driven by deep learning and the U-Net architecture. However, current datasets often focus on specific body regions or a limited number of organs, failing to capture the complex interactions within the entire human anatomy. Creating large, densely annotated datasets for full-body segmentation is a significant challenge due to the immense labor and expertise required from radiologists. This paper addresses the need for a comprehensive, whole-body anatomical dataset that can facilitate a more holistic understanding of human anatomy and its pathologies.
Methods
The authors developed an automated, expert-free pipeline for generating the DAP Atlas dataset. This process involves:
- Knowledge Aggregation: Combining fragmented anatomical knowledge from various publicly available datasets (e.g., Pediatric, Total Segmentator, SegThor, Amos, RibSeg, Verse) and private sources.
- nnU-Net Pseudo-Labeling: Utilizing the nnU-Net framework to train models on existing datasets and then applying these models to generate initial pseudo-labels for the target full-body CT scans.
- Anatomy-Guided Pseudo-Label Refinement: Employing a series of post-processing algorithms (e.g., Left-Right Split, Rib Counting, Non-largest Connected Component Suppression, Area Restriction, Sex-based Consistency) and rule-based knowledge derived from anatomical textbooks to refine and correct the pseudo-labels.
- Self-Training: A final iteration of self-training with a unified anatomical segmentation model to distill the aggregated knowledge into a single, efficient model and to ensure a seamless, high-resolution output.
Datasets
- Training/Knowledge Sources:
- AutoPET dataset (subset of 566 CT scans) - serves as the base for label aggregation.
- Pediatric dataset (359 CTs, 29 labels)
- Total Segmentator dataset (1024 CTs, 104 labels)
- SegThor dataset (60 thoracic CTs, 4 labels)
- CT50Abdomen dataset (50 abdominal CTs)
- MAL Cervix dataset (30 training, 20 testing abdominal CTs)
- Amos dataset (500 CTs, 15 abdominal organ categories)
- RibSeg dataset (490 CT scans)
- Verse dataset (374 CT scans, vertebra segmentation)
- ATM dataset (500 chest CT scans)
- PARSE dataset (203 CT images)
- Pelvic CT dataset (1184 CT images)
- Private body composition analysis model (yielding 9 labels)
- Private head-neck dataset (yielding 12 new labels)
- Newly Introduced Benchmark: The DAP Atlas dataset itself, comprising 533 full-body CT scans with 142 voxel-level anatomical labels.
Key Insights and Ideas
- Automated, Expert-Free Dataset Generation: The paper demonstrates that it's possible to create a large, densely annotated medical dataset without direct manual voxel-level annotation for every label, by cleverly aggregating existing knowledge and using rule-based refinement. This is crucial for scaling dataset creation in the medical domain.
- The Power of Knowledge Aggregation: By combining insights from multiple fragmented datasets, the authors were able to build a more comprehensive anatomical understanding than any single source could provide. This highlights the potential of leveraging diverse, existing resources.
- Anatomical Rules as a Crucial Refinement Step: The use of anatomical textbook knowledge and derived rules proved essential for correcting systematic errors and ensuring the plausibility of generated labels, especially for structures that are easily confused by models.
- Holistic Anatomy is Key: The paper strongly advocates for a holistic view of anatomy, emphasizing that structures interact and that understanding these interactions is vital for medical applications, moving beyond isolated organ segmentation.
Key Results
- DAP Atlas Dataset: A dataset of 533 whole-body CT scans with 142 anatomical labels, approved by human experts.
- Robust Segmentation Model (Atlas Prediction Model V2): Achieved an 85% Dice score on the BTCV abdomen dataset (an out-of-distribution test), which is on par with state-of-the-art supervised models. This model was developed using an adapted training strategy that improved robustness.
- Comparison with V1 Model: The robust V2 model showed a ~4% improvement in Dice score over the initial Atlas dataset model (V1), demonstrating the benefit of the adapted training.
- Qualitative Validation: Expert evaluation deemed the dataset "very impressive" for whole-body segmentation and useful for medical students and doctors. Anatomical insights derived from the dataset (e.g., volume vs. age plots) aligned with established medical knowledge.
Limitations and Future Work
- Limitations:
- While the dataset was expert-approved, the automated generation process means direct voxel-wise checks for every label are infeasible.
- The authors acknowledge that the initial model (V1) relied on post-processing, which is impractical in clinical settings.
- Some fine-grained structures (e.g., pulmonary artery branches) may be under-segmented, and the segmentation of the spinal canal could be more homogeneous. Borders of abutting organs might also show slight offsets.
- Future Work:
- The authors plan to make the dataset and trained models publicly available to foster further research.
- They suggest the dataset can serve as a cornerstone for investigating the relationship between anatomy and pathologies, such as how specific cancers affect different anatomical structures.
- The potential for future multi-modal (PET/CT) segmentation tasks is highlighted.
- The robust Atlas prediction model (V2) is intended for clinical environments where post-processing is not feasible.
Summary updated