TTA-MVUM: A Tumor and Text-Aware Multi-View Mamba Framework for Clinical Target Volume Auto-Segmentation in Nasopharyngeal Carcinoma Radiotherapy.

Authors

Abstract

Accurate segmentation of clinical target volumes (CTV) is critical for the success of radiotherapy in nasopharyngeal carcinoma (NPC). Manual contouring is time-consuming and prone to considerable inter-observer variability. Current auto-segmentation models rely exclusively on images, which may lead to sub-optimal accuracy. This study sought to substantially improve CTV auto-segmentation by incorporating tumor and clinical text information with multimodal deep learning. We retrospectively collected data from 184 NPC patients treated at our institution, with 144 cases allocated for model training and 40 for test. We propose a tumor- and text-aware multi-view UMamba (TTA-MVUM) model with text-image cross-attention modules to fuse visual and textual features. The model input comprised CT images, manually contoured primary gross tumor volume and metastatic lymph nodes, structured text contouring plan extracted by a large language model from electronic health records. Training was guided by a composite loss integrating Dice, cross-entropy and contrastive terms. Model performance was evaluated using Dice, 95% Hausdorff distance and averaged surface distance. The proposed TTA-MVUM model significantly outperformed state-of-the-art image-only and tumor-aware baselines, yielding mean Dice of 0.892 ± 0.035 for high-risk CTV, 0.738 ± 0.149 for low-risk CTV, and 0.889 ± 0.024 for the entire CTV, respectively, with similar improvements for other metrics. Using either tumor information or clinical text alone also surpassed baseline models, but performance remained inferior to the combined approach. Integrating tumor priors and text guidance can substantially improve the NPC CTV auto-segmentation accuracy and clinical applicability.

Tags

Topics