Deep Convolutional Neural Networks for Breast Cancer Detection

Abstract

Breast cancer is one of the main causes of cancer death worldwide. In the midst of the treatment of different disorders and diseases, one critical aspect in saving a patient is early detection. Correct detection and assessment of mammograms is hindered by human-error and inter-observer variations between pathologists. Existing convolutional neural network structures have shown promise in detection, but are hindered in their requirements for very large datasets to train on. The purpose of this paper is to explore a streamlined method classification of hematoxylin and eosin (H&E) stained tissue cancer mammograms into non-carcinomas and carcinomas using a small training set. This is done by the creation of more sample sets through changing elements of the data such as shear ratio and rotation. We assumed a 4-layer DCNN (deep convolutional neural network). We first train the DCNN with our augmented dataset, increasing dataset size by x200. We implement a highly accurate and reduced chance of overfitting gradient boosting algorithm. The overall classification accuracy of benign versus malignant was 88%.