What question did this study set out to answer?

To evaluate if the Pelvic U-Net can automate segmentation of pelvic organs-at-risk while maintaining accuracy and increasing efficiency.

March 7, 2026Open Access

Deep Learning Based 3D CT Image Segmentation of Pelvic Organs-at-Risk for Radiation Treatment Planning Using U-Net Architecture.

Key Points

To evaluate if the Pelvic U-Net can automate segmentation of pelvic organs-at-risk while maintaining accuracy and increasing efficiency.
Developed a 3D U-Net deep learning architecture for segmentation.
Trained using a dataset of 250 patients and evaluated on 50 independent test patients.
Applied geometric metrics, statistical analysis, and dosimetric comparisons for performance assessment.
Conducted a blinded Turing test to evaluate clinical acceptance.
Achieved Dice Similarity Coefficients (DSC) over 0.94 for high-contrast structures.
Demonstrated no significant difference in dose parameters between AI-generated and manual plans (P > 0.05).
94% of AI-generated contours were rated clinically acceptable in the Turing test.
Reduced contouring time from 45 minutes to around 8 minutes per patient, decreasing workload by 82.2%.

Abstract

Background and Purpose: The manual delineation of Organs-at-Risk (OARs) in radiotherapy planning is a labor-intensive process characterized by significant inter-observer variability and substantial time consumption. This study developed and clinically validated Pelvic U-Net, a 3D U-Net deep learning architecture designed to automate the segmentation of 13 pelvic OARs. The primary objective was to evaluate whether an AI-driven workflow could maintain clinical gold-standard accuracy while significantly increasing departmental efficiency. Materials and Methods: The model was trained and evaluated using a gender-diverse dataset of 250 patients using CT modality (160 male, 90 female). A robust preprocessing pipeline and an intensive 3D data augmentation strategy were implemented to handle anatomical variability. Model performance was rigorously assessed across 50 independent test patients using geometric metrics (DSC, IoU, HD95), statistical analysis (P-value, T-tests), and dosimetric comparisons (Dmean, Dmax, Vx). Furthermore, a clinical qualitative assessment was performed via a blinded Turing test, where a radiation oncologist graded 50 sets of contours on a 4-point Likert scale (Grade 1: Usable as-is to Grade 4: Unusable). Results: Geometric evaluation showed expert-level accuracy, with Dice Similarity Coefficients (DSC) exceeding 0.94 for high-contrast structures like the bladder and pelvic bones. Dosimetric validation confirmed that VMAT plans optimized on AI contours were clinically equivalent to manual plans, with no statistically significant differences (P > 0.05) in critical dose parameters. In the qualitative Turing test, 94% of the AI-generated contours were deemed clinically acceptable (56% Grade 1; 38% Grade 2), requiring only minor or no manual refinements. The automated workflow reduced active contouring time from 45 minutes to approximately 8 minutes per patient an 82.2% reduction in professional workload. Conclusion: The Pelvic U-Net serves as a robust and reliable supportive tool for radiation oncology. By providing high-fidelity automated contours that meet stringent clinical and dosimetric safety requirements, the model effectively standardizes the planning process and significantly accelerates clinical throughput. The results support the implementation of an "edit-and-approve" hybrid workflow to facilitate faster transitions from patient simulation to treatment delivery.

Bookmark

View Full Paper