Radiation Pneumonitis Prediction Using Dual-Modal Data Fusion Based on Med3D Transfer Network

Area of Science:

• Medical imaging
• Radiotherapy oncology
• Artificial intelligence in medicine

Background:

• Radiation pneumonitis (RP) is a lung inflammation following radiation therapy, with complex causes hindering accurate prediction.
• Developing reliable predictive models for RP is crucial for optimizing patient treatment and outcomes.

Purpose of the Study:

• To create a dual-modal prediction model for radiation pneumonitis (RP).
• To integrate pre-radiotherapy CT images with clinical and radiation dose data for enhanced RP prediction.

Main Methods:

• Utilized a Med3D transfer network with 3D channel attention for CT image feature extraction.
• Employed an autoencoder (AE) for deep feature dimensionality reduction.
• Applied univariate analysis and lasso regression for clinical and dose feature selection.
• Implemented adaptive feature fusion and a KAN classifier for binary classification.

Main Results:

• The dual-modal model achieved a precision of 74-76%, recall of 73%, and AUC of 86%.
• Performance significantly outperformed single-modality prediction approaches.
• Validated on a dataset of 117 chest cancer patients receiving radiotherapy.

Conclusions:

• Dual-modal data fusion effectively enhances radiation pneumonitis prediction accuracy.
• The proposed model offers a more reliable tool for identifying patients at risk of RP.
• This approach has the potential to improve radiotherapy planning and patient care.