MGTP-Seg: a mask guidance and text prompting network for gross tumor volume segmentation in esophageal cancer radiotherapy

Abstract

Objective. Existing automatic segmentation methods for esophageal cancer radiotherapy gross tumor volume rely solely on imaging features. Even when coarse masks are introduced as spatial priors, they still lack the pathological semantic guidance essential for clinical decision-making, resulting in segmentation outcomes that are misaligned with clinical practice. Approach. We propose a mask guidance and text prompting segmentation framework (MGTP-Seg) built upon the UNETR backbone. The mask guidance branch incorporates coarse segmentation masks generated by nnU-Net as spatial priors. The text prompting branch encodes clinical information including TNM stage and tumor location via a large language model, and aligns textual semantics with visual features through prompt tuning and a bidirectional attention mechanism. This design achieves end-to-end synergy among imaging details, spatial priors, and high-level clinical knowledge. Main Results. Multi-center experimental results demonstrate that MGTP-Seg achieves a Dice coefficient of 79.86% on the internal validation set, significantly outperforming 3D-UNet, HRNet, SAM-Med3D, and other mainstream methods. Ablation studies confirm the necessity of both branches. The model maintains stable performance on the external validation set (Dice: 78.76%), and text manipulation experiments reveal that the model dynamically adjusts segmentation in response to clinical descriptions, providing interpretability. Significance. MGTP-Seg not only offers an accurate, interpretable, and clinically aligned automatic delineation solution for esophageal cancer radiotherapy, but also establishes a novel methodological paradigm and empirical foundation for integrating spatial priors with semantic knowledge in medical image analysis.