Design of an Efficient Deep Neural Network for Multi-level Classification of Breast Cancer Histology Images

Abstract

Breast cancer is the second most prominent cause of death among women, and it is a major health problem across the world over the past many years. In this work, four hundred different biopsy images were collected from the Bio-imaging 2018 breast histology classification challenge by participating Grand Challenge on breast cancer histology images (BACH) for multiclass breast cancer histology image classification using deep learning technique. Computer-aided detection or diagnosis (CAD) system plays an important role in the detection and to increase death survival rate of women suffering from breast cancer, which can decrease death rate among women. The main purpose of this paper is to develop a CAD system which can be used to detect whether the input biopsy image fits one of the four different classes like benign, in situ, Invasive and Malignant. We have incorporated an efficient lightweight neural network such as MobileNet2.10ex instead of normal convolution neural network for feature extraction and at top a fully connected deep neural network is designed to do classification among different classes. The confusion matrix is developed in order to draw the model accuracy of both techniques. It is observed that MobleNet2.10ex is giving about 88.92 % training accuracy at 257 epochs and the corresponding loss value is 0.2249. In order to progress the performance during pre-processing, a staining technique was used to remove staining from a digital scan of histology microscopic image of fine-needle aspirate (FNA) slides and feature invariant approach such as data augmentation technique was used to rotate each image in pipeline about the axis 45° apart three times in order to increase dataset and prediction accuracy. The model was trained till 280 epochs to evaluate training and validation accuracy using Google Colab having strong GPU 2496 CUDA cores, 12 GB GDDR5 VRAM and 12.6 GB RAM.