ESTRO 2025 - Abstract Book

S1077

Clinical – Head & neck

ESTRO 2025

3263

Poster Discussion Performant Multimodal Nasopharyngeal Carcinoma Segmentation with 3D Gaussian-prompted Diffusion Deep Learning Model Jiarui ZHU, Jing Cai HTI, Hong Kong Polytechnic University, Hung hom, Kawloon, Hong Kong Purpose/Objective: Radiologists normally integrate multimodal information from CT and MRI series to operate a comprehensive Nasopharyngeal Carcinoma (NPC) gross tumor volume(GTV) delineation, but a performant multimodal segmentation deep learning model is still in absence. This study aims at developing a performant multimodal NPC GTV segmentation model with efficient multimodal information extraction based on 3D Gaussian Representing technique, and accurate clinical priors-guided step-wise segmentation based on diffusion Deep learning model.

Material/Methods:

784 NPC patients were retrospectively collected with CT, MRI series (t1,t1-contrast enhanced and t2) and clinical annotations, from four hospitals including Queen Mary’s Hospital ,Hong Kong(198 cases); Queen Elizabeth Hospital ,Hong Kong(196 cases); Xijing Hospital ,Xian, China(269 cases); and Western War Zone General Hospital ,Chengdu, China(121 cases). 70% of cases are selected as training set and 30% as validation set. All CT and MRI are cut into 2D slices and paired up with corresponding clinical GTV annotation masks. As can be seen in Figure 1(a), the image mask pairs are fed into a Denoising Diffusion Probabilistic Model (DDPM) for unconditional training, to encode the image-mask relations. Then as Figure 1(b) shows, based on clinical interests for NPC delineation and imaging modality characteristics, 3D Gaussian representing technique is employed to selectively extract key information from MRI-t1-CT, MRI-t2 and CT imaging individually. And finally, as can be seen from Figure1(c), the extracted 3D Gaussian representations which contain refined clinical information are used as conditions to prompt the Sampling process of DDPM and steply achieve a course-to-fine NPC GTV segmentation process.