Radiomics-Driven Machine Learning Models for Diagnosis of Pancreatic Adenocarcinoma

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Pancreatic adenocarcinoma is a highly lethal cancer with poor prognosis due to late diagnosis.
• Current diagnostic methods for pancreatic cancer lack sufficient sensitivity.
• Accurate and early diagnosis is critical for improving patient survival outcomes.

Purpose of the Study:

• To evaluate the effectiveness of radiomics features from Computed Tomography (CT) imaging for pancreatic adenocarcinoma detection.
• To assess the performance of machine learning models in identifying pancreatic cancer using radiomics data.
• To explore the potential of non-invasive diagnostic tools for pancreatic cancer.

Main Methods:

• Retrospective analysis of CT images from 100 participants (50 pancreatic adenocarcinoma, 50 controls).
• Extraction of radiomics features using 3D Slicer software.
• Application of Support Vector Machine (SVM), Logistic Regression (LR), and Random Forest (RF) classifiers with feature selection (LASSO, RFE, MI).

Main Results:

• The SVM classifier with LASSO feature selection achieved the highest performance (accuracy 0.83, AUC 0.89).
• Logistic Regression and Random Forest also showed strong results, with LASSO being the optimal feature selection method.
• SHAP analysis identified textural features (gray-level-non-uniformity, run-length-non-uniformity) as key diagnostic indicators.

Conclusions:

• Radiomics-based machine learning models demonstrate significant potential for enhancing pancreatic adenocarcinoma diagnosis.
• The combination of LASSO feature selection and classifiers like SVM, LR, and RF provides a robust framework.
• These findings suggest the development of accurate, non-invasive diagnostic tools for pancreatic cancer.