Deep-Learning Model Prediction of Radiation Pneumonitis Using Pretreatment Chest Computed Tomography and Clinical...

Area of Science:

• Medical Imaging and Radiation Oncology
• Artificial Intelligence in Healthcare
• Computational Biology and Bioinformatics

Background:

• Radiation therapy is a crucial treatment for lung cancer, but radiation pneumonitis is a significant limiting side effect.
• Existing models for predicting radiation pneumonitis lack reliability, necessitating the development of more accurate predictive tools.
• Accurate prediction of radiation pneumonitis is vital for optimizing treatment plans and improving patient outcomes.

Purpose of the Study:

• To develop and evaluate a comprehensive deep-learning model (MergeNet) for predicting radiation pneumonitis.
• To integrate diverse data sources, including chest computed tomography (CT), clinical, dosimetric, and laboratory data, for enhanced prediction accuracy.
• To compare the performance of MergeNet against traditional machine learning models like Support Vector Machine (SVM) and Light Gradient Boosting Machine (LGBM).

Main Methods:

• A retrospective analysis of 548 lung cancer patients treated between 2010 and 2021.
• Development of MergeNet, a deep-learning model combining a convolutional neural network with fully connected layers to process 3D CT, clinical, dosimetric, and laboratory data.
• Implementation and comparison of SVM, LGBM, and a convolution-only neural network using 3D Resnet-10 architecture and 4-fold cross-validation.

Main Results:

• MergeNet achieved an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.689, outperforming SVM (0.525), LGBM (0.541), and the convolution-only network (0.550).
• Statistical analysis (DeLong test) confirmed the significant superiority of MergeNet over SVM and LGBM (P < .001).
• The findings indicate that integrating multiple data types enhances the predictive capability for radiation pneumonitis.

Conclusions:

• The MergeNet model demonstrates superior performance in predicting radiation pneumonitis compared to existing models by integrating multimodal data.
• Despite its improved performance, MergeNet's predictive accuracy requires further enhancement before clinical implementation.
• This study highlights the potential of deep learning and multimodal data integration for advancing radiation pneumonitis prediction.