A study on breast cancer diagnosis based on deep convolutional neural networks

Breast cancer is a type of cancer that starts in the breast tissue. Symptoms of this cancer include a lump in the breast, a change in appearance, small skin indentations, discharge from the nipple, a sunken nipple, or red and dry skin (crusty) of the breast. In patients with metastasis, bone pain, swelling and enlargement of lymph nodes, shortness of breath or yellowness (jaundice) may also be seen. Risk factors for breast cancer include female sex, obesity, physical inactivity, alcoholism, hormone replacement therapy after menopause, ionizing radiation, early menarche (beginning of menstruation) at a younger age, late childbearing or no pregnancy, previous history of breast cancer. And there is a family history of breast cancer. About ۵-۱۰% of cases are genetic and inherited from parents, which can be mentioned BRCA۱ and BRCA۲ gene mutations. Most breast cancers originate from the cells of the wall of the milk ducts and milk-producing lobules. Those cancers that originate from the milk ducts are called "ductal carcinoma of the breast" and those that originate from the small milk-producing sacs are called "lobular carcinoma". Breast cancer detection based on the deep learning approach has gained much interest among other conventional-based CAD systems as the conventional based CAD system's accuracy results seems to be inadequate. The convolution neural network, a deep learning approach, has emerged as the most promising technique for detecting cancer in mammograms. In this paper we delve into some of the CNN classifiers used to detect breast cancer by classifying mammogram images into benign, cancer, or normal class. Our study evaluated the performance of various CNN architectures such as AlexNet, VGG۱۶, and ResNet۵۰ by training some of them from scratch and some using transfer learning with pre-trained weights. The above model classifiers are trained and tested using mammogram images from the mini-DDSM dataset which is publicly available. The medical dataset contains limited samples of data due to low patient volume; this can lead to overfitting issue, so to overcome this limitation data augmentation process is applied. Rotation and zooming techniques are applied to increase the data volume. The validation strategy used here is ۹۰:۱۰ ratio. AlexNet showed an accuracy of ۶۵ percent, whereas VGG۱۶ and ResNet۵۰ showed an accuracy of ۶۵% and ۶۱%, respectively when fine-tuned with pre-trained weights. VGG۱۶ performed significantly worse when trained from scratch, whereas AlexNet outperformed others. VGG۱۶ and ResNet۵۰ performed well when transfer learning was applied.

کلیدواژه ها:

Deep Learning ، Deep Convolution ، Neural Network ، Medical Imaging ، Mammograms (MGs)