Comparison of CNN and Transformer Models for Predicting the Effects of Anti-VEGF Drugs on Ovarian Cancer from Histopathology Images

Anti-vascular endothelial growth factor (anti-VEGF) therapy, such as Bevacizumab, treats colorectal, lung, kidney and breast cancer patients. In 2018, it was approved for treating ovarian cancer (OC) patients; however, when administered, it results in some adverse effects. Therefore, this therapy is given to only selected OC patients. Traditionally, the selection is done by manually examining the histopathology specimens of patients, which is time-consuming and vulnerable to inter-observer variability. An AI-based method could be beneficial in fixing these issues. In this study, we analyzed the suitability of popular AI-based models for predicting the effect of anti-VEFG therapy by analyzing patients’ histopathology images. We experimented with seven popular convolutional neural network (CNN) based models (VGG16, VGG19, ResNet50, InceptionNetV3, Xception, MobileNet and DenseNet12) and six transformer models (ViT-16, ViT-32, ViT-MaE, DiT, VAN and BEiT). We also investigated the role of image magnification in training the models and proposed a pipeline for automated patient selection leveraging the histopathology whole slide images. Our study finds that the transformer-based models achieve higher accuracy when trained with a sufficiently large dataset than the CNN-based models. The accuracy of the models was higher when trained using 20× images than 10×. The ViT-16 model achieved the highest accuracy (98.6%) and area under the curve score (AUC) (99.8%) for 20× images.