Histopathological Image Classification Using Ensemble Transfer Learning

Abstract

Histopathological image analysis of biopsy sample provides an accurate diagnosis method for cancer. Usually Pathologists examine the microscopic images of biopsy sample manually for the detection and grading of cancer. Automation in this field helps the pathologists to take a second opinion before confirming with their findings. We propose an effective method of automated cancer detection by combining the effect of transfer learning and ensemble learning. Six pre-trained models such as Densenet121, Resnet 50, Xception, EfficientNet B7, MobileNetV2, and VGG19 are used for preparing an ensemble model. A dataset contains 5547 H&E stained histopathological images of malignant and benign tissues are used to train and validate each models individually and obtained an accuracy of 77.9%, 79%, 79.8%, 78.3%, 77%, and 76% respectively. Based on the accuracy, best performing three models Resnet50, Xception, and EfficientnetB7 are selected to form an ensemble model. Then the layers and weights of these models are freezed and the output layers are concatenated to make an ensemble model. New dense layers are added to the ensembled model to provide a single output for binary classification. The model is compiled by an Adam optimizer with a learning rate of 0.001. The images are again applied to this ensemble model to classify the malignant and benign tissues and obtained an accuracy of 96% and precision of 96%.