Osama Mohamed et al., International Journal of Advanced Trends in Computer Science and Engineering, 10(1), January – February 2021, 122 – 132 122 ABSTRACT Breast cancer is one of the most serious diseases that affect women, so it must be discovered in the early stages to avoid complications such as redness of the skin, pain in the armpits or breast, and discharge from a nipple, possibly containing blood. Recently, the CAD system that is based on the classification of microscopic image play a vital rule to limit cancer disease and reduce cases. Microscopic image is the currently recommended image system used to detect cancer. A computer-aided diagnosis system will help radiologists to accurately detection of cancerous cells and achieve the best result. This paper proposes a deep learning technique that exploits CAD system features and microscopic images to fight breast cancer. The proposed technique builds a classification model based on the DenseNet-161 deep learning method. The proposed model classifies the microscopic images of breast cancer into benign with four types and malignant with four types. Our proposed technique is experimentally tested and the result confirmed that a proposed technique outperforms baseline techniques. Key words: Deep learning, Transfer learning, one fit cycle policy, Cyclical Learning Rates (CLRs), Breast cancer, Machine learning, Image classification 1. INTRODUCTION Automated detection of many types of breast cancer from a histopathological images is valuable for a clinical pathologists. According to reports raised in 2018, there are new cases of breast cancer which are around 2,088,849 over the world, this number represents 11.6% of all total cases of cancer types reported in this year, besides the number of death reported is 626,679 almost 6.6% from all total cases [1, 2]. Additionally, results in [3] report that incidence rates are reported as 19.3 per 100,000 women in Eastern Africa, to 89.7 per 100,000 women in Western Europe. It is expected that the number of new cases continues to grow until this number is increased to be 27 million in 2030. For the next 40 years, the diagnosis of breast cancer is based on X-ray, MRI (Magnetic Resonance Image) and ultrasound, etc. [4]. Pathologists diagnosis of the tissue slices by microscope examination [6]. Pathologists with no experience are likely to make diagnostic errors. The use of Computer-Aided Diagnosis (CAD) improves the diagnostic efficiency of histopathological image classification [6, 7]. A biopsy is known to be the only way to diagnose if the affected area is cancerous [5]. In detecting breast cancer through a classification of microscopic images, there are many challenges, the first of which is the histological images of breast cancer, complex, many colors, which are difficult to separate, to distinguish between cancer and other types, and the second is feature extraction is a big problem. Common methods for extracting image attributes are scale-invariant feature transform (SIFT) [12] and gray-level co-occurrence matrix (GLCM) [13]. Deep learning techniques have a large and distinct ability to extract image features and can quickly learn and is a solution [8, 9] to the problem of extracting image features, so it has been applied and achieved successful results in many areas such as the medical fields, the pharmaceutical industry, and other fields related to the field of computer vision [10, 11]. It is known that it is impossible to train deep learning from scratch by using a small set of data, so pre-trained models were used to save time that researchers obtained with training in big data called ImageNet. This is what is known as transfer learning, by re-training the fully connected layer of the model [16]. It is reported that deep learning achieves amazing success in many fields but it takes a long time is not necessary to train the model with the suboptimal hyperparameters so setting hyper-parameter is a big challenge [15] which face any researcher who uses deep learning and takes many years of experience to tune these parameters [14]. The solution is one cycle policy is a method that is used to reduce the time of training, improve performance, and tune all hyperparameters of deep learning models such as learning-rate, Weight-decay [14]. Figure 6 show that one cycle policy achieves training result higher than the result from the standard learning rate. 1.1 Contributions. The main contribution of this work can be summarized as follows: 1- We introduce a technique that classifies microscopic images into benign with four classes (Adenosis, Fibroadenoma, Tabular Adenoma, Phyllodes Tumors) and malignant with four classes (Ductal Carcinoma, Lobular Carcinoma, Mucinous Carcinoma, Papillary Carcinoma) Classification of breast cancer types based on deep learning approach Osama Mohamed 1 , Hala Nagy 2 , Hamdi Mahmoud 3 1 Faculty of computer and artificial intelligence, Beni-Suef University, Egypt, osama.mohamed@fcis.bsu.edu.eg 2 Faculty of computer and artificial intelligence, Helwan University, Egypt, hala.nagy@fci.helwan.edu.eg 3 Faculty of computer and artificial intelligence, Beni-Suef University, Egypt, Dr_hamdimahmoud@yahoo.co ISSN 2278-3091 Volume 10, No.1, January - February 2021 International Journal of Advanced Trends in Computer Science and Engineering Available Online at http://www.warse.org/IJATCSE/static/pdf/file/ijatcse171012021.pdf https://doi.org/10.30534/ijatcse/2021/171012021