Article Preview
Top1. Introduction
Breast Cancer (BC) is the primary reason behind the death of women globally and is answerable for numerous deaths every year (Afifi et al., 2019). BC cases are rising daily, and patients are increasing each day; health sciences are still troubled by the upper and proper prognosis of cancer and deep learning. Machine learning is attempting to supply technical advancements. As per the medical specialty reports, carcinoma breast cancer is hard to hunt out at the initial stage; the survival rate can increase if detected at the initial stage. Numerous methods and models have developed, and deep learning involvement gives rise to cancer's accurate and fast prognosis (Pacilè et al., 2020). There are several risk factors for cancer like increasing age, family history, biological time, internal secretion. BC has also reported at younger age women if having a primary child at an older age and women have never been pregnant. BC arises because the breast cell gets started growing unusually, and cells split quickly from the usual structure, frame a lump, and unfold to various parts (Yıldırım, 2018). Types of breast carcinoma like adenosis, fibroadenoma, phyllodes, cylindrical adenoma, mucinous carcinoma are shown in figure 1.
Figure 1. Four types of benign (first row) (a) adenosis (b) fibroadenoma (c) phyllodes tumor (d) tubular adenoma and malignant (second row) (a) ductal carcinoma (b) lobular carcinoma (c) mucinous carcinoma (d) papillary carcinoma tumor images from the Breast Cancer Histopathological Image Classification (BreakHis) dataset (Spanhol et al., 2016a).
Image classification has shown performance improvements using deep learning models (LeCun et al., 2015). The AlexNet model provides better results than conventional machine learning methods in detecting benign and malignant cancer cells (Spanhol et al., 2016b). State of the art has shown that deep learning models providing 83% accuracy (Bayramoglu N. et al., 2016), the pre-trained convolution neural network (CNN) model extracted the deep convolutional activation features from the network achieved 90% (Spanhol F. et al., 2017), and that BiCNN model up to 97% (Wei B. et al.,2017) accuracy in breast cancer detection. Deep learning still offers good precision with a large volume of data and shows its presence in detecting mitosis (Chen H. et al., 2016) and metastasis (Wang D. et al., 2016). The classification of histological images is very ambitious since invasive carcinoma and stroma have fragile borders, and the core element for the detection of cancer is tissue structures (Monaco J. et al., 2012) (Liu F.and Yang L., 2015). Hematoxylin and eosin (H&E) images show a variable appearance even with the same malignancy tissue biopsies (Giannakeas N. et al., 2017). Other issues with breast cancer detection with deep learning models are the availability of limited datasets, which cause overfitting (Pang et al., 2018) (Gavrilov et al., 2018); it also causes a lack of feature evaluation and results in cell imaging desegregation (Dapson et al., 2010). The VGG model uses VGG16 and a neural network to extract the properties (Gultom et al., 2018); the VGG16, VGG19, and ResNet50 models use a layered architecture but do not work well with smaller training datasets (Shallu & Mehra, 2018). AlexNet booths with large color images, LeNet5 has reduced training time and exceeds performance (Huang et al., 2015). The Inception v3 model based on deep learning was used to classify lung cancer and be a reliable model (Coudray et al., 2018). The contribution of this research is to design a model for the precise detection of nuclei, which remains a challenge as H&E images use the blue/purple color to represent the nuclei and the pink color to represent the cytoplasm very complicated (Veta et al., 2014). The main challenge is the presence of the maximum core in high-resolution images, noise, and variations in the core's texture, intensity, and shape (Fuyong Xing et al., 2014) (Chanho Jung & Changick Kim, 2010).