Related Experiment Video
Updated: Oct 29, 2025
08:17
Dynamic Lung Tumor Tracking for Stereotactic Ablative Body Radiation Therapy
Published on: June 7, 2015
15.9K
Deformation driven Seq2Seq longitudinal tumor and organs-at-risk prediction for radiotherapy
Donghoon Lee1, Sadegh R Alam1, Jue Jiang1
1Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, NY, USA.
Medical Physics
|July 10, 2021
Summary
This study introduces a novel 3D sequence-to-sequence model using deformation vector fields (DVFs) for predicting tumor and organ changes during radiotherapy. The model accurately forecasts anatomical deformations, improving adaptive radiotherapy outcomes.
Area of Science:
- Medical Imaging
- Radiotherapy Physics
- Computational Anatomy
Background:
- Radiotherapy requires accurate prediction of tumor and organ-at-risk (OAR) changes for effective treatment planning.
- Tumor inflammation and organ deformation pose significant challenges to longitudinal prediction accuracy.
- Adaptive radiotherapy necessitates flexible, rolling predictions grounded in pre-treatment imaging.
Purpose of the Study:
- To develop a novel 3D sequence-to-sequence model for predicting longitudinal anatomical changes in radiotherapy.
- To address the challenges of tumor inflammation, organ geometry changes, and flexible prediction requirements.
- To improve the accuracy and utility of predictions for adaptive radiotherapy.
Main Methods:
- A 3D sequence-to-sequence model utilizing Convolution Long Short-Term Memory (ConvLSTM) and deformation vector fields (DVFs) was developed.
- DVFs between timepoints and reference CTs were used to predict future anatomical deformations and volume changes.
- High-quality DVF training data were generated via hyper-parameter optimization using DICE coefficient and mutual information metrics.
- Model validation was performed on head-and-neck and non-small cell lung cancer radiotherapy datasets.
Main Results:
- The DVF representation and skip connections effectively addressed the blurring issue common in ConvLSTM predictions.
- The model achieved high DICE scores for predicting lung gross tumor volume (GTV) at weeks 4-6 (0.81-0.83).
- Accurate predictions were also obtained for post-treatment parotid glands (DICE scores of 0.81-0.85).
Conclusions:
- A novel DVF-based sequence-to-sequence model for medical image analysis was presented.
- The model leverages 3D imaging data for longitudinal GTV/OAR predictions in head-and-neck and lung radiotherapy.
- This approach shows potential for improving radiotherapy outcomes through enhanced adaptive treatment planning.
