Static-Dynamic coordinated Transformer for Tumor Longitudinal Growth Prediction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate prediction of tumor growth evolution is crucial for detailed clinical parameters and treatment planning.
• Existing longitudinal models often fail to capture complete tumor development due to loss of detailed growth information.

Purpose of the Study:

• To propose the Static-Dynamic coordinated Transformer (SDC-Transformer) for accurate longitudinal tumor growth prediction.
• To address limitations in existing models by extracting static features and dynamic growth associations.

Main Methods:

• Developed the SDC-Transformer model incorporating a Local Adaptive Transformer Module for pixel information and a Dynamic Growth Estimation Module.
• Utilized Enhanced Deformable Convolution and Cascade Self-Attention for enriched sampling and dynamic relationship analysis.
• Trained and tested the model on longitudinal tumor data from the National Lung Screening Trial (NLST) and Shanxi Provincial People's Hospital.

Main Results:

• The SDC-Transformer achieved high accuracy in longitudinal tumor prediction.
• Key performance metrics include RMSE (11.32), Dice (89.31%), Recall (90.57%), and Specificity (89.64%).

Conclusions:

• The SDC-Transformer model demonstrates effective and accurate longitudinal prediction of tumor growth.
• This technology can assist physicians in developing targeted treatment strategies and improving lung cancer diagnosis.