Abstract
Whole abdominal organ segmentation is critical for diagnosing lesions, radiotherapy planning, and follow-up. However, manual delineation of 3D volumes is time-consuming and expensive. We present the Whole abdominal ORgan Dataset (WORD), a large-scale benchmark containing 150 abdominal CT volumes (30,495 slices) with fine pixel-level and scribble-based annotations for 16 organs. We evaluate several state-of-the-art segmentation methods, investigate inference-efficient learning via knowledge distillation, and explore annotation-efficient learning using a novel intra-class intensity variance loss. Our results provide a new benchmark for abdominal multi-organ segmentation and highlight the remaining gaps between AI models and clinical practice.
Introduction: The Challenge of Abdominal Segmentation
Abdominal organ segmentation is a "high-effort" task in clinical workflows. Despite the potential of deep learning, progress has been hindered by two primary factors:
- Data Scarcity: Most public datasets are small or limited to a few organs (e.g., liver or kidneys). Many rely on semi-automatic labels that may carry model biases from the tools used to create them.
- Resource Intensity: High-resolution 3D CT scans demand significant GPU memory and long inference times. Furthermore, manual annotation of 16 organs in a single volume can take a senior oncologist over two hours, making large-scale data collection prohibitively expensive.
The WORD dataset aims to provide a "gold standard" to address these issues while exploring methods to reduce both computational and human labeling costs.
The WORD Dataset
The WORD dataset consists of 150 abdominal CT scans from a single center, featuring 16 organs: Liver, Spleen, Kidneys (L/R), Stomach, Gallbladder, Esophagus, Pancreas, Duodenum, Colon, Intestine, Adrenal, Rectum, Bladder, and Femur Heads (L/R).
To evaluate generalization, the study also utilizes external datasets including BTCV, TCIA, and re-annotated cases from LiTS to test how well models trained on WORD perform across different imaging protocols.
Methodology and Benchmarking
The paper is structured as a multi-faceted benchmark study:
- SOTA Benchmarking: Evaluation of CNN-based (nnUNet, ResUNet, UNet++) and Transformer-based (CoTr, UNETR) architectures.
- Inference-Efficient Learning: Using Knowledge Distillation (KD), a "heavy" teacher model (nnUNetV2) was used to train "lightweight" student models (e.g., ESPNet, LCOVNet), aiming for high accuracy with lower latency and fewer parameters.
- Annotation-Efficient Learning: Investigation of scribble-based supervision, where humans draw lines rather than coloring every pixel. To improve performance with sparse labels, the authors proposed an Intra-class Intensity Variance Minimization loss. This regularizer encourages pixels predicted as the same organ to have similar Hounsfield Unit (HU) intensities, helping the model "fill in" unannotated regions.
Key Insights and Results
CNNs vs. Transformers
Contrary to trends in general computer vision, CNN-based models—specifically nnUNetV2—consistently outperformed Transformer-based architectures on this 3D medical task, achieving a top mean Dice Similarity Coefficient (DSC) of 87.44%.
The "Solved" vs. "Unsolved" Divide
Large organs with distinct boundaries (liver, spleen, kidneys) are effectively "solved," with DSC scores exceeding 85%. However, small or hollow organs like the esophagus, duodenum, and adrenal glands remain challenging (DSC < 70%), indicating where future research should focus.
The Clinical Gap
By having oncologists revise AI predictions, the study quantified the "clinical gap." While AI is nearly ready for clinical use for large organs with minimal edits, it currently serves only as a "rough draft" for smaller, more complex structures.
Efficiency Gains
- Knowledge Distillation: Lightweight models like DMFNet saw significant DSC improvements (e.g., from 70.45% to nearly 80%) when guided by a teacher model.
- Scribble Success: The scribble-based method reached a mean DSC of 76.81%, outperforming the standard cross-entropy baseline (72.06%) and proving that intensity-based regularization is a powerful tool for weak supervision.
Conclusion
The WORD dataset provides a robust benchmark for abdominal segmentation. While current models excel at large organ segmentation, significant work remains for small, low-contrast organs and domain generalization across different hospital scanners. The proposed efficiency-focused methods demonstrate a viable path toward reducing the labor and hardware requirements for deploying AI in clinical settings.
Summary updated