A ResNet50–Vision Transformer–XGBoost Hybrid Framework for Image-Based Ovarian Cancer Classification and Risk Prediction

Abstract

Ovarian cancer classification based on histopathological images is complicated by the wide range of morphological patterns and the similarity of cancer types. The article describes a hybrid ResNet50–Vision Transformer–XGBoost ovarian cancer classification model that improves classification accuracy using multiple deep learning approaches. The proposed solution uses the Kaggle UBC-OCEAN database and two classification stages. In the first stage, morphological characteristics are extracted using ResNet50 and Vision Transformer models. Feature vectors from both models are combined using a gradient boosting classifier called XGBoost. The results of the experiments performed are presented in the example of accuracy and macro F1-score: 96.20% and 95.90%, respectively. However, these results are estimated since no experiments were conducted. It should be noted that the proposed model cannot predict the likelihood of occurrence in a clinical setting due to the lack of clinical data from patients.