Machine Learning Models Derived from [¹⁸F]FDG PET/CT for the Prediction of Recurrence in Patients with Thymomas

Area of Science:

• Oncology
• Radiology
• Medical Imaging

Background:

• Thymoma recurrence prediction is crucial for effective patient management.
• Preoperative imaging signatures may offer insights into recurrence risk.

Purpose of the Study:

• To develop and validate machine learning (ML) models for predicting recurrence in thymoma patients.
• To utilize radiomic and clinico-metabolic features from [18F]FDG PET/CT scans.

Main Methods:

• Retrospective analysis of 50 thymoma patients with preoperative [18F]FDG PET/CT.
• Extraction of 856 radiomic features (RFts) from PET and CT datasets following IBSI guidelines.
• Training and internal validation of ML models (Random Forest, Support-Vector-Machine, Tree) using selected RFts and clinico-metabolic signatures.

Main Results:

• Both CT-based and PET-based ML models effectively discriminated against recurrence-free survival (FFR).
• The CT model slightly outperformed the PET model, achieving an AUC of 0.970 with the Random Forest classifier.
• Key features included wavelet-based RFts from CT and metabolic parameters from PET.

Conclusions:

• ML models leveraging PET/CT radiomic features demonstrate significant potential for predicting thymoma recurrence.
• These findings suggest a future clinical application for personalized treatment strategies in thymoma management.