Survival prediction of ovarian serous carcinoma based on machine learning combined with pathological images and clinical information

Abstract

Ovarian serous carcinoma (OSC) has high mortality, making accurate prognostic evaluation vital for treatment selection. This study develops a three-year OSC survival prediction model using machine learning, integrating pathological image features with clinical data. First, a Convolutional Neural Network (CNN) was used to classify the unlabeled pathological images and determine whether they are OSC. Then, we proposed a multi-scale CNN combined with transformer model to extract features directly. The pathological image features were selected by Elastic-Net and then combined with clinical information. Survival prediction is performed using Support Vector Machine (SVM), Random Forest (RF), and XGBoost through cross-validation. For comparison, we segmented the tumor area as the region of interest (ROI) by U-net and used the same methods for survival prediction. The results indicated that (1) the CNN-based cancer classification yielded satisfactory results; (2) in survival prediction, the RF model demonstrated the best performance, followed by SVC, and XGBoost was less effective; (3) the segmented tumor ROIs are more accurate than those predicted directly from the original pathology images; and (4) predictions combining pathological images with clinical information were superior to those solely based on pathological image features. This research provides a foundation for the diagnosis of OSC and individualized treatment, affirming that both ROI extraction and clinical information inclusion enhance the accuracy of predictions.