Interpretable and Performant Multimodal Nasopharyngeal Carcinoma GTV Segmentation with Clinical Priors Guided...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate Gross Tumor Volume (GTV) segmentation is critical for image-guided radiation therapy (IGRT) in Nasopharyngeal Carcinoma (NPC).
• NPC GTV delineation is challenging due to complex infiltration and ambiguous boundaries, often requiring multimodal imaging (CT and MRI).
• Existing deep learning methods struggle with accurate multimodal segmentation, indicating issues in information extraction and integration.

Purpose of the Study:

• To develop a 3D Gaussian-prompted Diffusion Model (3DG-PDM) for enhanced GTV segmentation in NPC.
• To improve information extraction and multimodal integration for more precise NPC GTV delineation.
• To achieve clinically interpretable and accurate GTV segmentation for improved NPC treatment.

Main Methods:

• Proposed a 3D Gaussian-Prompted Diffusion Model (3DG-PDM) using 3D-Gaussian-Splatting for feature extraction from CT, MRI-t2, and MRI-t1-cefs.
• Developed a two-module model: Gaussian Initialization Module and Diffusion Segmentation Module for stepwise tumor contouring.
• Utilized a dataset of 600 NPC patients (480 training, 120 testing) with paired CT/MRI and GTV annotations.

Main Results:

• Achieved a Dice Similarity Coefficient (DSC) of 84.29% for primary GTV (GTVp) and 79.25% for metastatic GTV (GTVnd).
• Reported mean ASSD of 1.31 mm for GTVp and 1.19 mm for GTVnd, and HD95 of 4.76 mm for both.
• Demonstrated superior performance over five state-of-the-art methods and confirmed model interpretability through ablation studies.

Conclusions:

• The 3DG-PDM offers a performant and interpretable solution for multimodal GTV segmentation in NPC.
• This method significantly enhances precision in NPC GTV segmentation, aiding in improved radiation therapy.
• The study contributes a valuable tool for advancing precision medicine in NPC treatment planning.