Faster R-CNN Model for Predicting Endometrial Cancer Recurrence: A Five-Subclass Approach with Texture Analysis and Validation

Abstract

PurposeThis study investigates the prediction of endometrial cancer (EC) recurrence using MRI-based texture analysis and deep learning. EC recurrence remains a major clinical challenge, and accurate prognostic tools are needed. The study aims to identify significant imaging biomarkers for recurrence risk and evaluate the performance of a Faster R-CNN deep learning model in classifying recurrence-free survival (RFS) and recurrence (REC) groups. Materials and MethodsA total of 112 patients with histopathologically confirmed EC who underwent pre-treatment MRI between 2009 and 2018 were included. MRI texture analysis was performed using histogram and gray-level co-occurrence matrix (GLCM) methods, analyzing perimeter, integrated density, entropy, and homogeneity. A Faster R-CNN deep learning model was trained and tested on five image subclasses, using 20,000 iterations to optimize performance. REC and RFS were classified based on clinical outcomes, and probability scores were calculated to define the REC score. Feature importance was assessed using a Random Forest model. ResultsSignificant clinical predictors of recurrence included high-grade tumors, FIGO stage III-IV, and cervical invasion. Texture analysis revealed notable differences in perimeter, integrated density, contrast, entropy, and homogeneity between REC and RFS groups. Faster R-CNN achieved an average precision of 0.86 and an AUC of 0.61. Feature importance analysis identified area as the most significant factor, while REC_score was independent of texture-based parameters. ConclusionMRI texture analysis combined with deep learning provides valuable predictive insights for EC recurrence. The study highlights the potential of integrating machine learning with imaging biomarkers for personalized risk assessment, offering a promising approach for improving EC patient management..