A-Eval [Paper]

🌟 Highlights

• 📊 A benchmark focused on cross-dataset generalizability in abdominal multi-organ segmentation.
• 🧠 In-depth analysis on model generalizability across different data usage scenarios and the role of model size.

📚 Datasets

We train models on the official sets of FLARE22, AMOS, WORD, and TotalSegmentator, and evaluate them using their official validation sets as well as BTCV's official training set.

Note: While these datasets do have test sets, FLARE22, AMOS, and BTCV do not make their test labels publicly available. Therefore, for consistent evaluation, we use validation sets instead of test sets in A-Eval, regardless of label availability.

Dataset	Modality	# Train	# Test	# Organs	Region
FLARE22	CT	50 labeled 2000 unlabeled	50	13	North American European
AMOS	CT & MR	200 CT 40 MR	100 CT 20 MR	15	Asian
WORD	CT	100	20	16	Asian
TotalSegmentator	CT	1082	57	104	European
BTCV	CT	-	30	13	North American
A-Eval Totals	CT & MR	1432 labeled CT 2000 unlabeled CT 40 MR	257 CT 20 MR	8	North American European Asian

To ensure a meaningful and fair comparison across datasets, we evaluate the models’ performance based on a set of eight organ classes shared by all five datasets. We unify these labels using an overlapped label system. The corresponding code for label systems and label conversion can be found in the repository: label_systems.py and convert_label_2_overlap_label.py.

Organ Class	FLARE22	AMOS	WORD	TotalSegmentator	BTCV	A-Eval
Liver	✓	✓	✓	✓	✓	✓
Kidney Right	✓	✓	✓	✓	✓	✓
Kidney Left	✓	✓	✓	✓	✓	✓
Spleen	✓	✓	✓	✓	✓	✓
Pancreas	✓	✓	✓	✓	✓	✓
Aorta	✓	✓	✗	✓	✓	✗
Inferior Vena Cava	✓	✓	✗	✓	✓	✗
Adrenal Gland Right	✓	✓	✗	✓	✓	✗
Adrenal Gland Left	✓	✓	✗	✓	✓	✗
Gallbladder	✓	✓	✓	✓	✓	✓
Esophagus	✓	✓	✓	✓	✓	✓
Stomach	✓	✓	✓	✓	✓	✓
Duodenum	✓	✓	✓	✓	✗	✗

🏆 Results

💡 DSC

Train/Test	FLARE22	AMOS CT	WORD	TotalSeg	BTCV	CT Mean	AMOS MR	All Mean
FLARE22 w/o PL	89.20	76.53	85.94	74.06	86.11	82.37	24.77	72.77
FLARE22 w/ PL	91.98	87.53	87.15	85.55	87.35	87.91	42.74	80.38
AMOS CT	89.14	93.02	89.01	86.39	86.84	88.88	70.08	85.75
AMOS MR	61.47	73.97	45.30	48.08	77.60	61.28	91.73	66.36
AMOS CT+MR	89.81	93.24	89.36	88.42	87.66	89.70	92.72	90.20
WORD	86.86	87.53	90.92	80.58	84.69	86.12	27.38	76.33
TotalSeg	90.32	89.65	86.30	95.12	87.73	89.82	38.72	81.31
Joint Train	91.98	92.42	88.88	93.87	88.90	91.21	90.87	91.15

💡 NSD

Train/Test	FLARE22	AMOS CT	WORD	TotalSeg	BTCV	CT Mean	AMOS MR	All Mean
FLARE22 w/o PL	90.19	80.25	90.76	76.56	89.28	85.41	23.96	75.17
FLARE22 w/ PL	93.46	90.92	92.01	88.29	90.94	91.12	44.19	83.30
AMOS CT	89.49	96.47	94.82	89.28	91.65	92.34	72.92	89.11
AMOS MR	59.97	48.69	43.93	48.09	61.61	52.26	95.22	59.42
AMOS CT+MR	90.46	96.80	95.18	91.36	92.53	93.27	96.58	93.82
WORD	88.73	92.34	95.75	83.47	88.74	89.81	30.75	79.96
TotalSeg	91.96	94.02	92.46	97.33	92.72	93.70	40.44	84.82
Joint Train	93.58	96.46	95.28	96.10	93.80	95.04	95.28	95.08

💡 Visualization

🎫 License

This project is released under the Apache 2.0 license.

🙏 Acknowledgement

• Special thanks go to the creators and maintainers of the public datasets that made our research possible:
  • FLARE22
  • AMOS
  • WORD
  • TotalSegmentator
  • BTCV
• Thanks to the SOTA framework of: nnUNet

👋 Hiring & Global Collaboration

• Hiring: We are hiring researchers, engineers, and interns in General Vision Group, Shanghai AI Lab. If you are interested in Medical Foundation Models and General Medical AI, including designing benchmark datasets, general models, evaluation systems, and efficient tools, please contact us.
• Global Collaboration: We're on a mission to redefine medical research, aiming for a more universally adaptable model. Our passionate team is delving into foundational healthcare models, promoting the development of the medical community. Collaborate with us to increase competitiveness, reduce risk, and expand markets.
• Contact: Junjun He(hejunjun@pjlab.org.cn), Jin Ye(yejin@pjlab.org.cn), and Tianbin Li (litianbin@pjlab.org.cn).