Source-linked AI summary
Improving Clinical Target Volume Segmentation Accuracy using Anatomical Priors and Active Learning for the AGITG TOPGEAR Clinical Trial
Phillip Chlap, Mark Lee, Trevor Leong, Matthew Field, Jason Dowling, Hang Min, Julie Chu, Jennifer Tan, Phillip K. Tran, Tomas Kron, Annette Haworth, Martin A. Ebert, Shalini K. Vinod, Lois Holloway
TL;DR
Limited, resource-intensive curated datasets make training medical image segmentation models difficult, especially for the anatomically complex TOPGEAR CTV. This study evaluates anatomical priors and active learning for sequential model development, finding that their combination achieved the highest overall accuracy.
Problem
Limited curated medical-imaging datasets make segmentation-model training difficult, while TOPGEAR’s CTV requires landmarks across complex abdominal anatomy for automated contour QA.
Method
The study adds TotalSegmentator-derived anatomical priors and uses hybrid confidence-and-DSC active learning to select cases for sequential retraining.
Results
Anatomical priors and active learning each improved CTV segmentation accuracy, while their combination achieved the highest overall accuracy.
Takeaways & Limitations
These techniques support segmentation-model development for automated contour QA in prospective radiotherapy clinical trials.
Takeaways & Limitations
TotalSegmentator was not developed specifically for radiotherapy, so task-specific radiotherapy models may be more suitable anatomical-prior sources.
Abstract
from arXiv · showhide
Training deep learning-based medical image segmentation models is challenging with limited curated datasets. For AGITG TOPGEAR, a gastric cancer trial, the Clinical Target Volume (CTV) is complex and defined by multiple anatomical landmarks, making upfront training data preparation difficult for an automated contour QA segmentation model. We investigate anatomical priors, derived from surrounding organ segmentations, to provide spatial context and improve TOPGEAR CTV segmentation accuracy. We also evaluate active learning, iteratively expanding the training dataset by selecting cases expected to improve performance. One hundred TOPGEAR CT scans were retrospectively analyzed. An initial set of 10 expert-contoured cases was used to train an nnU-Net model. TotalSegmentator generated a voxel-wise anatomical prior map from surrounding structures as an additional input channel. Active learning was simulated over four iterations, selecting cases by model uncertainty and segmentation performance. All models used five-fold cross-validation for an ensemble uncertainty measure. Evaluation used a hold-out testing set of 50 cases. The anatomical prior improved CTV segmentation accuracy, increasing mean Dice Similarity Coefficient (DSC) from 0.84 to 0.86. Active learning similarly improved performance to 0.86, with greatest benefit in the final round. Combining the anatomical prior with active learning achieved the highest accuracy, with a DSC of 0.87. Model uncertainty correlated with DSC, supporting its use in identifying suboptimal predictions and guiding active learning. Anatomical priors and active learning each improved CTV segmentation accuracy and generalizability, with their combination achieving the best performance, supporting integration into segmentation model development for automated contour QA in radiotherapy clinical trials.
Introduction
Limited, resource-intensive curation makes training accurate and generalizable medical-image segmentation models difficult, especially for the anatomically complex TOPGEAR CTV. This study evaluates anatomical priors and active learning as strategies for improving segmentation with limited initial data.
- Large, curated datasets are difficult and resource-intensive to obtain for medical-image segmentation.
- The TOPGEAR CTV is defined by multiple abdominal anatomical landmarks, complicating preparation of representative training cases and gold-standard contours.
- Small training datasets can perform sub-optimally because they provide limited anatomical and imaging diversity.
- Data augmentation may improve performance but remains limited in extending generalization beyond characteristics represented in the original training data.
- Anatomical priors from surrounding-organ segmentations can provide voxel-wise spatial context for target volumes defined by their relationships to adjacent structures.
- Active learning iteratively adds cases selected using model predictions, often prioritizing uncertain examples that may represent underrepresented anatomy or imaging characteristics.
- The study simulates sequential case availability to evaluate anatomical priors and active learning for improving TOPGEAR CTV segmentation in future automated contour QA workflows.
Method
The study trained nnU-Net CTV segmentation models using expert-contoured cases, five-fold ensembles, anatomical priors, and simulated active learning. Cases were selected using confidence and DSC, with evaluation on a reserved hold-out set.
- Dataset: 100 planning CT scans were divided into an initial 10-case training set, a 40-case trial set, and a 50-case hold-out testing set.The initial contours came from five radiation oncologists and were consolidated using a consensus workshop and STAPLE reference segmentation.
- Dataset: The esophageal-extension scenario was excluded from the initial annotations and removed consistently from trial and testing reference contours.The excluded region concerned tumor extending 4cm into the esophagus; connected components overlapping the esophagus and smaller than 50mm² were removed.
- Model training: nnU-Net v2.5.1 3D full-resolution models used five-fold cross-validation, with each ensemble member selected by validation performance.Each fold used 80% of the data for training and 20% for validation, and the final ensemble averaged fold predictions before voxel-wise maximum-probability classification.
- Anatomical Prior: TotalSegmentator produced a normalized single-channel map of 117 anatomical structures, which was concatenated with CT as a two-channel model input.The anatomical prior supplied spatial context from structures relevant to CTV definition or surrounding anatomy, and models with and without it were compared.
- Active Learning: Active learning simulated four rounds of sequential case availability, selecting cases flagged by low confidence or low DSC for retraining.Updated models were trained from scratch on the full expanded dataset, with and without anatomical priors; selection parameters were chosen empirically.
Results
Anatomical priors and active learning each improved TOPGEAR CTV segmentation, with their combination achieving the strongest reported performance. Confidence scores were positively correlated with DSC, while anatomical priors added modest computational overhead.
- Anatomical priors: 0.84±0.09 to 0.86±0.05: adding anatomical priors significantly improved mean DSC and reduced variability.The paired t-test found significant improvement across all metrics (p<0.05).
- Active learning: 0.87±0.03: the final active-learning model with anatomical priors achieved the best overall DSC.Without anatomical priors, active learning reached 0.86±0.05; four rounds were needed to match the anatomical-prior-only result of 0.86.
- Active learning: Active learning without anatomical priors produced incremental DSC gains, while early-round improvements were minimal and not statistically significant.The largest benefit occurred in the final active-learning round.
- Confidence and performance: Confidence score 𝒞 showed a consistent positive Pearson correlation with DSC across models.The figure also marks inter-observer variability at 0.89 ± 0.02 as an approximate upper bound for model performance.
- Computational cost: 22 hours versus 18 hours: anatomical priors increased approximate single-fold training time, while inference increased from 6 to 7 minutes per image.Generating each TotalSegmentator anatomical-prior map required approximately 6 minutes per image.
Discussion
Anatomical priors and active learning improved TOPGEAR CTV segmentation in complementary ways, while their combination achieved the strongest reported performance. The discussion also identifies practical scope boundaries and directions for prospective clinical-trial deployment.
- Performance: DSC reached 0.87 with the combined model, compared with an inter-observer variability DSC of 0.89 that may represent an upper performance bound.The IoV estimate came from a limited and potentially less diverse subset of the dataset.
- Deployment: Confidence scores correlated with DSC and could help identify unreliable predictions for manual review in automated contour QA.Low-confidence cases may reflect underrepresented anatomy, poor image quality, or artefacts.
- Anatomical priors: Anatomical priors improved coverage of the duodenum and stomach, including contrast-enhanced cases absent from the initial training set.The prior supplied explicit localization of surrounding structures, supporting spatial context when imaging characteristics differed from training data.
- Applicability: The approach is especially suited to anatomically defined targets when datasets are small, anatomy is underrepresented, and boundaries lack strong intensity gradients.The TOPGEAR CTV is defined relative to nearby anatomical landmarks.
- Active learning: Active learning selected additional training cases using model confidence and DSC, but significant improvement emerged only in the final round.Intermediate models performed worse in some regions, reflecting case variability and sequential data availability.
- Limitations: Important limitations include empirically chosen active-learning parameters, no benchmarking against alternative methods, and TotalSegmentator’s lack of radiotherapy-specific development.Future work could evaluate task-specific anatomical-prior sources and additional active-learning rounds.
Conclusion
Anatomical priors and active learning each improved TOPGEAR CTV segmentation accuracy, while their combination achieved the highest overall accuracy and supports automated contour QA development.
- Anatomical priors and active learning each improved CTV segmentation accuracy, with their combination achieving the highest overall accuracy.These findings support their use in model development for automated contour QA in prospective radiotherapy clinical trials.