Breast Cancer Detection from Histopathological Biopsy Images Using Transfer Learning

Breast cancer is one of the most common deadly diseases among women around the world. Early detection of the disease improves the cure rate for patients. In this paper, a dataset named VBCan is introduced, which is composed of images of hematoxylin and eosin (H&E) stained lymph node sections, collected from two specialized hospitals in Vietnam. The dataset has 3529 images at resolution of 512x512. Then, a two-stage method is introduced to evaluate the accuracy of the breast cancer detection, i.e. a combination of feature extraction by one of state-of-the-art CNNs namely VGG-16, GoogLeNet or ResNet-50 and various conventional machine learning classifiers. In addition, two transfer learning techniques are applied as firstly we fine-tune the networks by updating parameters of all layers on a benchmark dataset namely Patch Camelyon 2017 by initializing their weights from the pretrained models on ImageNet. Then, the features are extracted from the CNNs to train different machine learning classifiers on VBCan dataset to improve the deep learning models. The optimization results show that the topmost accuracy on VBCan is 96.98% by using ResNet-50 model to extract features and Softmax as a classifier. It is also recognized that the recall of VGG-16 model using its original classifier is the highest value of 97.76% among other models while GoogLeNet model achieves the highest precision of 98.58%.