2113 - Guided by the Guidelines: Clinical Guideline-Driven Refinement of Deep Learning for Pancreatic Tumor Segmentation

Purpose/Objective(s): Pancreatic cancer's aggressive nature and poor prognosis necessitate precise CT tumor segmentation for effective radiotherapy planning. However, challenges such as complex anatomy, tumor variability, and inter-operator inconsistency hinder both manual and automated methods. Current purely data-driven deep learning models often fail to capture critical tumor-vessel relationships emphasized in clinical guidelines. This study introduces a novel framework integrates these guidelines, focusing on the Tumor-Vessel Interface (TVI), to enhance segmentation accuracy and consistency, potentially standardizing results across clinical settings.

Materials/Methods: Our guideline-driven, iterative refinement framework addresses these challenges. Initial tumor/OAR segmentation is refined using a Vision-Language Large Model (VLLM), QWen2.5-VL-70B to analyze tumor-vessel spatial relationships. The VLLM takes as input both the initial segmentation and textual descriptions from clinical consensus guidelines (Australian Gastrointestinal Clinical Research Collaboration Group & Trans Tasman Radiation Oncology Group TVI definition), outputting textual descriptions that capture clinically relevant spatial relationships. These descriptions are then embedded using ClinicalBERT, and the resulting text-derived, clinically enriched embeddings are used as additional input to the segmentation network, emphasizing the tumor-vessel interface during iterative refinement. This refinement repeats three times. Evaluation used 63 pancreatic cancer contrast-enhanced helical CTs (1.5mm slices, target volumes 11.3-71.1 cc, 80/20 split, expert annotations, predominantly head tumors).

Results: Our method achieved superior mean DSC (Dice Similarity Coefficient) 0.68 ± 0.04, HD95 (Hausdorff Distance at 95th percentile) 7.5 ± 1.5 mm, Sensitivity 0.75 ± 0.05, Specificity 0.92 ± 0.02, and low Relative Volume Error 8.0% ± 3.0% on the test set, outperforming VISTA 3D (p<0.05). Compared to VISTA 3D, our guideline-informed approach showed an 8% relative DSC improvement and 15% HD95 reduction, demonstrating clinical guidance benefits.

Conclusion: Our VLLM-guided, guideline-driven deep learning method significantly improves CT pancreatic tumor segmentation accuracy and boundary delineation. By explicitly modeling crucial tumor-vessel relationships and perivascular tissue as recommended clinically, our approach addresses key limitations, enhancing clinically relevant tumor segmentations for radiotherapy and disease management. However, the computational complexity associated with using a VLLM and performing multiple iterations of refinement may limit real-time clinical applicability. Future work will focus on optimizing the model for faster inference while maintaining accuracy.