Novel Neural Network for Breast Cancer Diagnosis

Abstract

The second biggest cause of mortality for women worldwide is Breast Cancer (BC). BC diagnosis by hand using histological breast pictures is expensive, time-consuming, and non-generalizable. Using a CNN to directly learn features from entire slide images is an alternative way for feature extraction. A significant number of labelled images, which can occasionally be challenging to get, are necessary for training the CNN. Reusing a pre-trained CNN model for feature attainment with huge image datasets from other disciplines is the solution. The BreakHis dataset contains images of BC histology, and in this article, we provide a “Novel CNN” architecture using Transfer Learning for identifying those images. This model's binary classification-benign and malignant-allows it to quickly and accurately diagnose breast cancer. In the suggested framework, DenseNet-201 pre-trained model is used to attain features from the histopathological pictures. Then, to generate a reliable hybrid model, the attained features are applied into the Global Average Pooling Layer, followed by Dropout, Batch-Normalization, and Dense Layers. The proposed model had a 99.75% accuracy rate. These encouraging findings will open the door to utilize this model as an automated tool to help clinicians diagnose breast cancer and may improve the survival rate for the disease.