Joint Distance-Aware and Multi-Organ Constrained Learning for Automated CTV Segmentation In Cervical Cancer Radiotherapy
Abstract
Purpose
This study aimed to develop an accurate and automated deep learning model for segmenting the clinical target volume (CTV) in cervical cancer radiotherapy, addressing the limitations of time-consuming and variable manual contouring.
Methods
We propose JDCC-Net, an innovative two-stage architecture. The first stage performs efficient 3D global context modeling using a U-Net enhanced with xLSTM blocks, which capture long-range spatial dependencies across the CT volume with linear computational complexity. This stage generates a coarse initial segmentation. The second stage conducts a 2.5D local refinement, processing the target slice along with its neighbors. It incorporates OAR distance maps as prior knowledge and employs dedicated Cross-Slice and In-Slice Attention modules to integrate contextual information. Crucially, this stage is trained with a novel distance-constrained loss function, explicitly enforcing the model to learn and adhere to clinically plausible geometric distances between the predicted CTV and surrounding OARs.
Results
Evaluated on a clinical dataset of 323 cervical cancer patients, JDCC-Net achieved a Dice Similarity Coefficient of 0.878 and a 95% Hausdorff Distance of 12.31 mm, significantly outperforming seven state-of-the-art segmentation methods across all metrics, including boundary accuracy (Average Surface Distance: 3.15 mm). The framework demonstrated exceptional robustness in complex cases, such as those with retroperitoneal metastases, and showed strong volume consistency with manual contours (Concordance Correlation Coefficient =0.908). Ablation studies quantitatively confirmed the indispensable contribution of each proposed component the two-stage design, the distance maps, and the distance-constrained loss to the final performance.
Conclusion
JDCC-Net provides an accurate, efficient, and clinically informed solution for CTV segmentation, with the potential to standardize radiotherapy planning and reduce clinical workload. The integration of explicit geometric constraints represents a significant step toward more trustworthy and adoptable AI in clinical practice.