2209 - An Adaptive 3D Large Vision Model for the Auto-Delineation of Clinical Target Volume and Organs at Risk in Pancreatic Cancer Radiotherapy

Purpose/Objective(s): Accurate delineation of clinical target volumes (CTV), gross tumor volumes (GTV), and organs at risk (OARs) remains a time-consuming challenge in pancreatic cancer radiotherapy due to complex retroperitoneal anatomy. This study proposes Adapted-MedSAM3D, a transformer-based 3D segmentation framework with adaptive spatial attention mechanisms, to address critical limitations in existing automated solutions for multi-structure radiotherapy planning.

Materials/Methods: A cohort of 120 pancreatic adenocarcinoma patients with contrast-enhanced CT simulation scans (1mm slice thickness) was analyzed. All ground truth contours were established through multi-expert consensus (3 senior radiation oncologists) adhering to RTOG 0822 guidelines. The architecture integrates hybrid convolution-transformer blocks with organ-specific prior encoding, trained using combined Dice and boundary-focused loss functions. Performance was rigorously evaluated against UNet3D, DPNUnet, and SAM-Med 3D using DSC, HD95.

Results: The DSC values obtained by Adapted - MedSAM 3D for the test data were 86.6±5.1%, 85.8±4.1% for CTV, GTV; 94.9±1.1%, 75.3±10.9%, 79.2±5.0%, 82.1±6.1%, 94.6±1.3%, 95.5±1.1%, 90.9±3.0% for the liver, duodenum, spinal cord, bowel, left kidney, right kidney, stomach. The HD95% values for CTV and GTV were 11.9±4.1mm, 10.1±3.8mm respectively; The HD95% values for the liver, duodenum, spinal cord, bowel, left kidney, right kidney, stomach were 9.2±2.8mm, 14.6±3.8mm, 2.6±0.6mm, 12.9±4.1mm, 3.7±1.1mm, 3.3±1.2mm, 10.5±4.0mm, respectively. Compared to the baseline network, the results of the clinical target show significant improvement (GTV: DSC 7.5%?,HD95% 2.2mm?, CTV: DSC 7.1%?, HD95% 2.5mm?).

Conclusion: The proposed Adapted - MedSAM 3D framework achieves clinically acceptable accuracy for multi - structure pancreatic radiotherapy planning, demonstrating significant improvements in anatomically challenging regions and could significantly reduce the radiation oncologists’contouring time.