Machine learning for predicting accuracy of lung and liver tumor motion tracking using radiomic features

Area of Science:

• Medical Imaging
• Radiomics
• Machine Learning

Background:

• Internal tumor motion prediction often relies on external respiratory signals, but the correlation is complex and patient-specific.
• Accurate tumor motion tracking is crucial for effective radiotherapy, especially for lung and liver cancers.
• Existing methods face challenges due to the intricate relationship between internal tumor movement and external surrogates.

Purpose of the Study:

• To develop and evaluate radiomic models for predicting the accuracy of tumor motion tracking using external surrogates.
• To identify reliable algorithms for accurate and stable tumor motion tracking in lung and liver cancer patients.
• To leverage computed tomography (CT) radiomic features for improved internal/external motion correlation.

Main Methods:

• Extracted radiomic features from CT images of 108 lung and 71 liver cancer patients.
• Utilized cross-validation least absolute shrinkage and selection operator (LassoCV) for feature selection.
• Trained and tested 26 machine learning models (13 for lung, 13 for liver) to classify tracking accuracy based on fitting errors > 2 mm.

Main Results:

• Selected 33 radiomic features for lung and 22 for liver cancer.
• Achieved high Area Under the Curve (AUC) values, up to 0.941 for lung (SVC, logistic regression) and 0.892 for liver (logistic regression).
• Multilayer perceptron (MLP) and Wide & Deep (W&D) models demonstrated superior performance and stability in both lung and liver cancer tracking.

Conclusions:

• Respiratory-sensitive radiomic features from CT images effectively capture internal/external motion relationships.
• Developed a rapid and accurate radiomics-based method for classifying tumor tracking accuracy.
• MLP and W&D algorithms show excellent potential for reliable lung and liver tumor motion tracking.