Abstract

Esophageal cancer, a serious malignancy, is primarily treated with radiotherapy (RT). Accurate delineation of gross tumor volume (GTV) and clinical target volume (CTV) is essential, yet manual contouring is time-consuming and automatic methods remain challenging. Adaptive radiotherapy enables personalized target adjustments, but tumor changes following initial RT complicate delineation, and automatic second-phase replanning is underexplored. We propose the Unspecified Pretraining and Modular Adaptation (UPMA) framework for context-aware second-phase esophageal target delineation. UPMA pretrains its network on diverse datasets to comprehend tumor and anatomical features, then modularly adapts it for second-phase tasks with minimal parameters. The method improves Dice scores by over 23% for GTV and over 5% for CTV, while reducing learnable parameters by over 73% during adaptation. In clinical evaluations, 73.7% and 52.6% of the predicted GTVs and CTVs required only minor edits. UPMA generalizes across diverse patient characteristics and imaging factors, supporting automated adaptive radiotherapy for precise, personalized treatment.

Fig. 1 Tumor morphological and volume changes after first-phase radiotherapy. a GTV is the tumor area, and CTV includes the GTV and subclinical disease without a specific boundary. Both GTV and CTV rely on relevant anatomical structures. b The volume distributions of GTV and CTV differ between the first and second phases of adaptive radiotherapy, with a reduction of both volumes in the second phase. Failure to promptly update target volumes may reduce treatment effectiveness and increase the risk of damage to surrounding healthy tissues. c Our treatment is a two-phase adaptive radiotherapy with systematic replanning between phases. The proposed method allows personalized updates of target volumes during treatment based on the patient’s actual response and tumor changes, enabling more precise dose delivery and improved organ-at-risk protection. d–f The proposed Unspecified Pretraining and Modular Adaptation paradigm (UPMA). We perform unspecified pretraining on the network and adapt it to context-aware second-phase tasks with minimal parameters, enabling the integration of various priors. This approach enhances the automation of replanning and adaptive updates during second-phase treatment, for better tumor control and greater protection of organs at risk. Graphic elements adapted from UXWing (https://uxwing.com/license/). Plot generated using Matplotlib (https://matplotlib.org/stable/project/license.html). 3D visualization rendered using 3D Slicer (https://slicer.readthedocs.io/en/latest/user_guide/about.html).

Fig. 2 Segmentation performance gap between the two RT phases and the superiority of UPMA for GTV delineation in adaptive radiotherapy. a A performance gap between two phases is illustrated, which indicates increased difficulty in esophageal GTV segmentation for the second phase. UPMA outperforms all SOTA methods and achieves superior results in the more challenging second-phase RT scenario, facilitating better outcomes in multi-phase adaptive radiotherapy. Circle center positions denote DSC and HD-95. Circle size represents RVD, with gray circles at the bottom left corresponding to RVD=25, 40, and 55. b UPMA reduces over-prediction by effectively handling tumor changes. SOTA methods perform reasonably well in delineating the first-phase GTV (pink arrows) but exhibit poor performance in the second-phase GTV. Orange arrows show false positives in the second phase, indicating that SOTA methods neglect tumor volume changes after first-phase RT. c UPMA also more precisely delineates second-phase GTV in the lower esophagus that can not be detected by SOTA methods (blue arrows). b, c UPMA achieves superior performance by reducing over-prediction and missed tumor volumes, which improves tumor control and OAR protection across multiple phases of adaptive radiotherapy. The color bars indicate the absolute surface distance (mm), showing the 3D differences between predictions and ground truth masks. Plot generated using Matplotlib (https://matplotlib.org/stable/project/license.html). 3D visualization rendered using 3D Slicer (https://slicer.readthedocs.io/en/latest/user_guide/about.html).

https://doi.org/10.1038/s44387-026-00143-w