Predictive model of Ki67 expression level in osteosarcoma based on weakly supervised segmentation and multi-type feature fusion

Abstract

Background and objective

Osteosarcoma is a highly malignant bone tumor that occurs primarily in children and adolescents. Ki67 protein expression level (detected through immunohistochemistry) is an important indicator for assessing tumor proliferative activity. This study aims to develop an efficient and low-cost artificial intelligence model to predict Ki67 expression levels from pathological images.

Methods

73 hematoxylin and eosin-stained (H&E) whole slide images (WSIs) of osteosarcoma specimens were analyzed. Tumor regions were segmented using weakly supervised learning, followed by extraction of 215 nuclear features including shape, texture, spatial and topological features through the Hover-Net network. Feature selection was performed using five methods: least absolute shrinkage and selection operator (LASSO), mutual information (MI), recursive feature elimination (RFE), Wilcoxon rank sum test (WRST), and extreme gradient boosting (XGBoost), with the top 5 features selected from each method. These features were subsequently integrated with 8 machine learning classifiers: adaptive boosting (AdaBoost), balanced random forest (BalancedRF), k-nearest neighbors (KNN), light gradient boosting machine (LightGBM), multilayer perceptron (MLP), quadratic discriminant analysis (QDA), random forest (RF), and support vector machine (SVM) to determine the optimal hybrid model.

Results

By combining 5 key features with 8 machine learning classifiers, we selected the optimal hybrid model (XGBoost+SVM). This model demonstrated the best performance in accuracy (0.767 ± 0.018), recall (0.872 ± 0.036), F1-score (0.800 ± 0.012), and receiver operating characteristic-area under curve (ROC-AUC) (0.884 ± 0.045). The model showed both high accuracy and high sensitivity in Ki67 detection.

Conclusion

Our model provides an automated and reliable solution for osteosarcoma Ki67 assessment, reducing dependence on traditional immunohistochemistry. Its excellent performance indicates strong potential for clinical translation.