Breast Cancer Detection from Histopathological Images using Deep Learning and Transfer Learning

Microscopic analysis of breast tissues is one of the most precise technique used for the diagnosis of breast cancer. However, this technique is time consuming as it requires scanning of various breast tissue images with different magnification levels. Computer assisted diagnosis techniques that make use of machine learning methods can automate some of the tasks in the breast cancer diagnostic process. This paper conducts a comparative study of the use of CNN and transfer learning based DenseNet121 deep learning models for automatic classification of breast histopathological images into malignant or benign. This study uses 7909 histopathological images of benign and malignant types. The impact of image magnification, scaling and rotation on the model's accuracy was investigated. Results show DenseNet121 outperformed (86.6% accuracy) for the 100X magnification level with 128*128 image size scaling. Transfer learning significantly increased the training accuracy by 16.4% for 100X magnification level.