Explainable breast cancer detection from Histopathology images using transfer learning and XAI

Abstract

Nowadays, research on many diseases such as cancer has been ongoing to determine how to reduce and minimise the effect. There are many characteristics of cancer that can be identified by their consistent cell proliferation and unique subgroups. Among cancer, breast cancer is responsible for many deaths each year and early detection increases the chance of survival. The proposed method employs three base models, VGG19, ResNet50V2 and MobileNetV2 which are trained on the BreakHis dataset, a public dataset of breast histopathological images. Furthermore, technology such as CNN and ML have become a tool for cancer researchers to identify cancer cells more efficiently. Feature extractors such as MobileNetV2, ResNet50V2 etc. models have been used for classification and detection. MobileNetV2 is a feature extractor for segmentation and object detection. Nearly all of the latest AI technology uses ResNet to build cutting-edge systems. A well-liked method for producing a class specific heatmap using a trained CNN, a specific input image and a class of interest is called Grad-CAM. We trained our model using the transfer learning techniques using MobileNetV2, ResNet50V2, VGG19 as the base model and the weights of Im ageNet. The model had an accuracy rate of 94.86%, 94.38%, 95.65% respectively. The features extracted from the last layer of the trained models are fused using concatenation and ensemble methods to improve the performance of the classifiers. Several linear classifiers including K-Nearest Neighbors (KNN), Stochastic Gradient Descent (SGD), AdaBoost, XGBoost, Decision Tree and Random Forest are used to classify the fused features. The results of the experiments show that the proposed method achieved high accuracy, with KNN classifier achieving the best result of 97.535% and Random Forest classifier achieving 97.455%. The proposed method is effective in breast cancer prediction and can assist pathologists in the diagnosis of breast cancer.