Applying a random projection algorithm to optimize machine learning model for predicting peritoneal metastasis in...

Area of Science:

• Radiomics and Machine Learning in Oncology
• Medical Imaging Analysis
• Computational Pathology

Background:

• Accurate prediction of gastric cancer metastasis is crucial for treatment planning, particularly for neoadjuvant chemotherapy.
• Computed tomography (CT) imaging offers potential for non-invasive assessment of cancer spread.
• Small and imbalanced datasets pose challenges for developing robust machine learning models.

Purpose of the Study:

• To investigate the efficacy of a random projection algorithm in optimizing a radiomics-based machine learning model.
• To predict the risk of peritoneal metastasis in gastric cancer patients using CT image data.
• To evaluate the performance of machine learning models with different feature selection techniques.

Main Methods:

• Retrospective analysis of CT images from 159 gastric cancer patients (121 with, 38 without peritoneal metastasis).
• Segmentation of primary gastric tumors and extraction of 315 radiomic features using a computer-aided detection scheme.
• Development of five Gradient Boosting Machine (GBM) models, each employing a different feature selection method (including random projection) and a synthetic minority oversampling technique, trained with leave-one-case-out cross-validation.

Main Results:

• The GBM model incorporating the random projection algorithm achieved a significantly higher prediction accuracy (71.2%) compared to other models (p<0.05).
• The optimal model demonstrated a precision of 65.78%, sensitivity of 43.10%, and specificity of 87.12% for predicting peritoneal metastasis.
• Feature selection using random projection proved superior in enhancing the predictive performance of the radiomics model.

Conclusions:

• CT-derived radiomic features of primary gastric tumors contain valuable information for predicting peritoneal metastasis.
• The random projection algorithm is a promising technique for generating optimal feature vectors, thereby improving machine learning model performance.
• This approach facilitates non-invasive risk stratification for gastric cancer patients, aiding in treatment decisions.