Deep Transfer Learning with CNN Models for Breast Cancer Prediction Using Histopathological Image Classification

Abstract

Breast cancer (BCr) is still a major problem that affects a lot of people all over the globe. Quick medical action and assistance may be provided, if possible, dangers of breast cancer are identified in a timely manner. Here, we provide new approaches to breast cancer diagnosis that make use of TL and Convolutional Neural Network (CNN) architecture. The CNN architectures that have already been trained, include InceptionV3, xception, Googlenet, and Inceptionresnetv2, all of which used the BeakHis dataset with different magnifications. While the TL part uses pretrained weights to build the model, the CNN part does a masterful job of extracting spatial characteristics from histopathology images. We use inceptionresnetv2 to address issues caused by a lack of labeled data and to avoid overfitting. To enable the network to learn relevant characteristics from histopathology pictures, our CNN component is initialised using a pre-trained CNN model using large image datasets. The proposed Inceptionresnetv2 model identifies possible instances of BCr with an impressive recall (83.03%), accuracy (87.02%), precision (86.71%), and F1-score (86.35%) for 40x magnification. The utilised Grad-CAM, Heat maps, and Occlusion Sensitivity (OS) tools to analyse and visualise model decisions. These methods draw attention to the areas where the CNN makes predictions. In the long run, this strategy has the ability to improve women's health care by assisting doctors with early risk assessment and intervention for BCr.