Main Session
Sep
28
PQA 03 - Digital Health Innovation and Informatics, Patient Safety & Quality, and Radiation and Cancer Biology
Presenter(s)
István Megyeri, PhD - GE HealthCare, Wauwatosa, WI
V. Gila1, A. Marinovszki1, M. Mian2, L. Ferenczi3, and I. Megyeri1; 1GE Healthcare, Hungary, Szeged, Hungary, 2GE Healthcare, Wauwatosa, WI, 3GE Healthcare, Hungary, Budapest, Hungary
Purpose/Objective(s):
Weakly supervised segmentation methods can reduce annotation burden, a key limitation for developing robust AI systems in medical imaging. We evaluated a practical weak-supervision strategy that uses annotations from only a subset of slices. We hypothesized that the required annotation proportion varies by organ due to differences in shape complexity and variability. Our goal was to identify organ-specific annotation requirements that maintain high segmentation accuracy while minimizing labeling effort.Materials/Methods:
We performed abdominal organ segmentation using the MRI subset of the AMOS dataset, including 40 training and 20 testing 3D volumes with 13 annotated organs (aorta, duodenum, esophagus, gallbladder, bilateral adrenal glands, bilateral kidneys, liver, pancreas, postcava, spleen, and stomach). Two fully supervised baselines (nnU-Net and the MultiTalent Network) were trained using full-slice annotations. Weak supervision was evaluated using the MultiTalent Network with annotated-slice ratios of 20%, 40%, 60%, and 80%. Additionally, an organ-specific strategy was tested, assigning each organ the minimum annotation percentage required to achieve 99% of its fully supervised Dice score. Performance was assessed using Dice and surface-distance metrics. We post-processed the predictions by keeping only the largest 3D connected components of the predicted mask.Results:
Segmentation accuracy improved with increasing annotation ratios and reached a performance plateau near 80%. Several organs (aorta, kidneys, liver, spleen, stomach) required only ~20% of annotated slices to achieve near–fully supervised accuracy, whereas the gallbladder and postcava required 40% annotated slices. The esophagus and pancreas required 60%, and the duodenum and adrenal glands required 80% or more annotation. The organ-specific allocation reduced the mean annotation requirement to 41% without meaningful loss in Dice or surface accuracy relative to full supervision.Conclusion:
Weak annotations can substantially reduce labeling effort for abdominal MRI segmentation while maintaining high accuracy for most organs. Organ-specific annotation strategies are particularly effective, reducing the required annotation load by more than half. For the studied organs, only 41% of annotated slices were needed to achieve near-fully supervised performance. These findings support scalable dataset creation and offer practical guidance for efficient clinical AI development.| Training annot. | Test | ||||
| Network | Organ specific | % of Annot. Slices | Dice | HD95 | ASSD |
| MultiTalent | N | 20 | 0.8415 | 5.28 | 1.24 |
| N | 40 | 0.8584 | 4.38 | 0.84 | |
| N | 60 | 0.8648 | 4.64 | 0.84 | |
| N | 80 | 0.8694 | 4.47 | 0.82 | |
| Y | 41 | 0.8666 | 4.46 | 0.82 | |
| N | 100 | 0.8716 | 4.29 | 0.83 | |
| nnU-Net | N | 0.8641 | 3.87 | 0.80 | |