DART-Net: A Novel Deep Learning Framework for Precise Radiotherapy Planning with Automated Multiorgan Segmentation and RTSTRUCT Generation.

PURPOSE: Manual delineation of pelvic organs remains a critical bottleneck in radiotherapy planning, consuming valuable clinical time and introducing interobserver variability. While existing deep learning approaches have shown promise, they struggle with complex anatomical boundaries and lack seamless integration into clinical workflows. This study addresses these limitations by introducing Dual-Attention Residual Technology Network (DART-Net), the first framework to unify dual-encoder architecture, attention mechanisms, and residual connections for automated segmentation and direct RTSTRUCT generation.

METHODS

Our novel architecture processes both raw and normalized computed tomography (CT) volumes simultaneously through parallel encoders, enabling superior feature extraction. The integration of attention mechanisms with residual connections, a combination previously unexplored in pelvic segmentation, allows precise focus on clinically relevant regions when preserving gradient flow. The model was trained on 125 expert-annotated pelvic CT scans with comprehensive data augmentation to ensure anatomical variability representation.

RESULTS

DART-Net achieved state-of-the-art performance with significant improvements over existing methods: Dice Similarity Coefficient scores of 0.965 for bladder (vs. 0.94 with prior methods), 0.908 for prostate, and 0.840 for rectum, with Hausdorff Distance values between 1.4 and 2.5 mm, representing an 18.6% reduction in bladder boundary error compared to the best previous approach. Expert validation by radiologists and radiation oncologists confirmed clinical acceptability of the auto-generated contours. The model uniquely automated bridges the research-practice gap through RTSTRUCT file generation, enabling seamless integration with commercial treatment planning systems.

CONCLUSION

DART-Net establishes a new benchmark for artificial intelligence-assisted radiotherapy planning by harmonizing technical innovation with clinical practicality. By reducing contouring time and improving anatomical precision, this framework addresses critical workflow inefficiencies in radiation oncology when potentially enhancing treatment outcomes through more consistent organ delineation.