Class Probability-guided Ensemble Learning-based Semantic Segmentation to Predict Cancerous Regions on Hematoxylin and Eosin-stained Images

Ensemble-based machine learning aims to obtain generalized prediction outcomes improving predictive performance by combining multiple neural network models. The success of ensemble-based machine learning in providing accurate predicted regions has been widely presented in many research areas, but it is rarely investigated in the field of histopathology image analysis. In this paper, we propose a class probability-guided ensemble learning method that aims to enhance the effectiveness of semantic segmentation on cancer-related regions in histopathological images. The proposed method combines the prediction probabilities of ensemble-based semantic segmentation with the prediction probabilities of weighted ensemble-based image classification. Ensemble-based semantic segmentation is conducted by averaging from five backbone networks, ResNetl8,ResNet50, Mobilenetv2, Xception, and InceptionResNetv2 from the DeepLabv3 model. Weighted ensemble-based image classification is conducted by computing the weighted averages from five different semantic segmentation models: GoogLeNet, AlexNet, InceptionResNet, VGG16, and ResNet50. The performance of the proposed method was evaluated and comprehensively analyzed using the BCSS dataset including cancer-related hematoxylin and eosin images and immune-related hematoxylin and eosin images. The results show that the proposed methods outperform the state-of-the-art semantic segmentation models in terms of accuracy, F1 measure, and IoU evaluation metrics.