Combining outputs from multiple deep learning auto-segmentation systems can generate more accurate and robust organ contours than individual models alone, according to a multi-institutional proof-of-concept study. While automated contouring tools are increasingly adopted to improve the consistency and efficiency of clinical workflows, no single deep learning system performs optimally across every patient case, anatomical organ, and image slice. Because individual systems occasionally generate poor-quality or outlier contours, researchers evaluated a multi-system fusion framework designed to leverage their complementary strengths.
The investigation utilized 35 abdominal computed tomography datasets gathered from three participating institutions to examine contour delineation for four specific anatomical structures: the stomach, duodenum, liver, and right kidney. To generate the underlying automated contours, the research team deployed three commercial deep learning auto-segmentation systems alongside one non-commercial research model. The resulting outputs served as the baseline inputs for testing systematic fusion strategies against verified clinical reference contours.
Investigators evaluated two distinct fusion strategies across the multi-system dataset. The first approach applied conventional Simultaneous Truth and Performance Level Estimation, known as STAPLE, which merged all available automated contours directly. The second method introduced a selecting-and-combining pipeline termed scDLAS. In the scDLAS framework, previously developed contour quality classification models used supervised machine learning and seven quantitative metrics to categorize each contour slice as acceptable, requiring minor edits, or requiring major edits before fusion.
Within the scDLAS pipeline, only contour slices placed into the highest available quality category were retained for subsequent STAPLE fusion. Researchers tested various scDLAS configurations that incorporated differing numbers of deep learning systems. Performance evaluation against the reference contours relied on the Dice similarity coefficient, the 95th-percentile Hausdorff distance measured in millimeters, and shifts in the overall distribution of contour-quality categories across the anatomical datasets.
Analysis of the individual systems confirmed that no single deep learning model consistently achieved the best contouring performance across all four examined organs. Conventional STAPLE generally matched or exceeded the performance of individual systems while maintaining comparable or lower interpatient variability. Specifically, STAPLE achieved mean Dice similarity coefficients of 0.88 for the stomach, 0.70 for the duodenum, 0.95 for the liver, and 0.92 for the right kidney, with corresponding 95th-percentile Hausdorff distances of 10.8 mm, 18.0 mm, 5.5 mm, and 6.8 mm.
The best-performing scDLAS configuration further improved contour precision across all target structures, raising the mean Dice similarity coefficient to 0.90 for the stomach, 0.74 for the duodenum, 0.96 for the liver, and 0.94 for the right kidney. The highest duodenal Dice score was produced by a three-system configuration, whereas a four-system configuration achieved or tied the top score for the remaining organs. The researchers concluded that the framework demonstrates the feasibility of combining multiple systems to supply more dependable initial contours for subsequent clinical review and editing.
Subject of Research: Deep learning auto-segmentation fusion for abdominal CT contouring
Article Title: Enhancing auto-segmentation accuracy and robustness through multiple deep learning systems
Article References: Zhang, Y., Zarenia, M., Sarosiek, C., Yuan, L., Paulson, E., & Qi, X. S. (2026). Enhancing auto-segmentation accuracy and robustness through multiple deep learning systems. PLOS One, 21(11), e0363053. https://doi.org/10.1371/journal.pone.0363053
Image Credits: AI Generated
DOI: 10.1371/journal.pone.0363053
Keywords: deep learning auto-segmentation, computed tomography, medical image segmentation, STAPLE fusion, contour quality classification, Dice similarity coefficient, Hausdorff distance, abdominal imaging, stomach contouring, duodenum segmentation, liver segmentation, kidney contouring, supervised machine learning, radiation oncology, clinical contouring, multi-institutional study
Cite Scienmag News
Blake Davidson. (November 21, 2026). Multi-System Fusion Improves Deep Learning Auto-Segmentation for Abdominal Organs. Scienmag. https://scienmag.com/multi-system-fusion-improves-deep-learning-auto-segmentation-for-abdominal-organs/
Blake Davidson. "Multi-System Fusion Improves Deep Learning Auto-Segmentation for Abdominal Organs." Scienmag, 21 November 2026, https://scienmag.com/multi-system-fusion-improves-deep-learning-auto-segmentation-for-abdominal-organs/. Accessed 21 November 2026.
Blake Davidson. "Multi-System Fusion Improves Deep Learning Auto-Segmentation for Abdominal Organs." Scienmag. November 21, 2026. https://scienmag.com/multi-system-fusion-improves-deep-learning-auto-segmentation-for-abdominal-organs/








