PET radiomics-based lymphovascular invasion prediction in lung cancer using multiple segmentation and multi-machine...

Area of Science:

• Oncology
• Radiology
• Data Science

Background:

• Lymphovascular invasion (LVI) is a critical prognostic factor in non-small cell lung cancer (NSCLC).
• Accurate prediction of LVI is essential for tailoring treatment strategies and improving patient outcomes.
• Positron Emission Tomography (PET) radiomics offers a non-invasive approach to extract quantitative imaging features for predictive modeling.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting LVI in NSCLC patients using multi-segmentation PET radiomics.
• To compare the performance of various segmentation techniques, feature selection methods, and classification algorithms.
• To identify the optimal combination of methods for accurate LVI prediction.

Main Methods:

• One hundred twenty-six NSCLC patients were included.
• Multiple PET image segmentation methods (LAC, FCM, KM, Watershed, RG, IT) were applied.
• 105 radiomic features were extracted and analyzed using MRMR, RFE, and Boruta feature selection.
• Classifiers including MLP, LR, XGB, NB, and RF were employed, with and without SMOTE for data balancing.

Main Results:

• The combination of SMOTE, Iterative Thresholding (IT) segmentation (45% threshold), RFE feature selection, and Logistic Regression (LR) classifier achieved the highest performance (AUC=0.93, ACC=0.84, SEN=0.85, SPE=0.84).
• Another model using SMOTE, Fuzzy-C-mean (FCM) segmentation, MRMR feature selection, and LR classifier also showed strong results (AUC=0.92, ACC=0.87, SEN=1, SPE=0.84).
• The highest accuracy (ACC=0.9) was achieved with IT segmentation (45%/50% thresholds), Boruta feature selection, and Naive Bayes classifier without SMOTE.

Conclusions:

• PET radiomics combined with machine learning can accurately predict LVI in NSCLC patients.
• The choice of segmentation method and machine learning algorithm significantly impacts prediction performance.
• These findings support the potential of radiomics-based prediction for personalized NSCLC treatment strategies.