Enhancing auto-segmentation accuracy and robustness through multiple deep learning systems

Purpose Deep learning auto-segmentation (DLAS) can improve the efficiency and consistency of clinical contouring; however, no single DLAS system performs optimally across all cases, organs, and image slices, and each system may produce occasional poor-quality or outlier contours. Because multiple DLAS systems are available in some institutions, this proof-of-concept study evaluated a multi-system fusion framework designed to leverage their complementary strengths and generate more accurate and robust auto-contours, with potential extension to consensus generation from multiple manual contours. Methods Using 35 abdominal CT datasets from three institutions, contours of the stomach, duodenum, liver, and right kidney were generated by three commercial DLAS systems and one non-commercial research DLAS model. Two fusion approaches were evaluated: conventional Simultaneous Truth and Performance Level Estimation (STAPLE), which combined all available DLAS contours, and a selecting-and-combining DLAS (scDLAS) approach. For scDLAS, previously developed contour quality classification (CQC) models used supervised machine learning and seven quantitative metrics to classify each contour slice as acceptable, requiring minor edits, or requiring major edits. Only contours in the highest available quality category were retained for STAPLE fusion. scDLAS configurations incorporating different numbers of DLAS systems were evaluated against the reference contours using the Dice similarity coefficient (DSC), 95th-percentile Hausdorff distance (HD95), and changes in the distribution of contour-quality categories. Results No individual DLAS system consistently achieved the best performance across all four organs. STAPLE generally matched or outperformed the individual systems while maintaining comparable or lower interpatient variability. For the stomach, duodenum, liver, and right kidney, respectively, STAPLE achieved mean ± SD DSC values of 0.88 ± 0.10, 0.70 ± 0.13, 0.95 ± 0.02, and 0.92 ± 0.04, with corresponding HD95 values of 10.8 ± 12.6, 18.0 ± 12.3, 5.5 ± 2.2, and 6.8 ± 9.6 mm. The best-performing scDLAS configuration increased the mean DSC to 0.90 ± 0.09, 0.74 ± 0.14, 0.96 ± 0.02, and 0.94 ± 0.03, respectively. The highest duodenal DSC was achieved by 3scDLAS, whereas 4scDLAS achieved or tied the highest DSC for the other organs. Conclusion The proposed framework demonstrates the feasibility of leveraging complementary outputs from multiple DLAS systems to improve the accuracy and robustness of auto-segmented contours. These findings support further evaluation of multi-system fusion as a strategy for generating more reliable initial contours for subsequent clinical review and editing.

Authors

Institutions

Publication Details

Journal
PLoS ONE
Published
2026-11-19
DOI
https://doi.org/10.1371/journal.pone.0363053
Primary Topic
Medical Image Segmentation Techniques
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Enhancing auto-segmentation accuracy and robustness through multiple deep learning systems

E.S. Paulson, Christina Sarosiek, M. Zarenia, Ying Zhang et al.
PLoS ONE
Medical Image Segmentation Techniques
article

Enhancing auto-segmentation accuracy and robustness through multiple deep learning systems

E.S. Paulson, Christina Sarosiek, M. Zarenia, Ying Zhang, Lulin Yuan, X. Sharon Qi
article en

Abstract

Purpose Deep learning auto-segmentation (DLAS) can improve the efficiency and consistency of clinical contouring; however, no single DLAS system performs optimally across all cases, organs, and image slices, and each system may produce occasional poor-quality or outlier contours. Because multiple DLAS systems are available in some institutions, this proof-of-concept study evaluated a multi-system fusion framework designed to leverage their complementary strengths and generate more accurate and robust auto-contours, with potential extension to consensus generation from multiple manual contours. Methods Using 35 abdominal CT datasets from three institutions, contours of the stomach, duodenum, liver, and right kidney were generated by three commercial DLAS systems and one non-commercial research DLAS model. Two fusion approaches were evaluated: conventional Simultaneous Truth and Performance Level Estimation (STAPLE), which combined all available DLAS contours, and a selecting-and-combining DLAS (scDLAS) approach. For scDLAS, previously developed contour quality classification (CQC) models used supervised machine learning and seven quantitative metrics to classify each contour slice as acceptable, requiring minor edits, or requiring major edits. Only contours in the highest available quality category were retained for STAPLE fusion. scDLAS configurations incorporating different numbers of DLAS systems were evaluated against the reference contours using the Dice similarity coefficient (DSC), 95th-percentile Hausdorff distance (HD95), and changes in the distribution of contour-quality categories. Results No individual DLAS system consistently achieved the best performance across all four organs. STAPLE generally matched or outperformed the individual systems while maintaining comparable or lower interpatient variability. For the stomach, duodenum, liver, and right kidney, respectively, STAPLE achieved mean ± SD DSC values of 0.88 ± 0.10, 0.70 ± 0.13, 0.95 ± 0.02, and 0.92 ± 0.04, with corresponding HD95 values of 10.8 ± 12.6, 18.0 ± 12.3, 5.5 ± 2.2, and 6.8 ± 9.6 mm. The best-performing scDLAS configuration increased the mean DSC to 0.90 ± 0.09, 0.74 ± 0.14, 0.96 ± 0.02, and 0.94 ± 0.03, respectively. The highest duodenal DSC was achieved by 3scDLAS, whereas 4scDLAS achieved or tied the highest DSC for the other organs. Conclusion The proposed framework demonstrates the feasibility of leveraging complementary outputs from multiple DLAS systems to improve the accuracy and robustness of auto-segmented contours. These findings support further evaluation of multi-system fusion as a strategy for generating more reliable initial contours for subsequent clinical review and editing.

PLoS ONEVol. 21(11)
University of California, Los Angeles (US), Virginia Commonwealth University (US), Medical College of Wisconsin (US), MedStar Georgetown University Hospital (US), The University of Texas Southwestern Medical Center (US)
Medical Image Segmentation Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon