br Introduction br Breast cancer is still one of the
Breast cancer is still one of the top leading causes of death in women worldwide . To diagnose a wide variety of breast cancer types properly, it is necessary to apply a medical test (commonly performed by a surgeon), followed by a microscopic analysis of breast tissue. In the first stage of this process, the doctor has to cut section biopsy materials and then stain them using hematoxylin and eosin staining. The hematoxylin solution binds deoxyribonucleic Nigericin (DNA) and highlights nuclei, while eosin binds proteins and highlights other structures . In the second stage of this analysis, pathol-ogists evaluate tissue biopsies by visualizing highlighted regions in digitized images using microscopes. The evaluation of tissue biopsies allows the identification of early clues of tissue biopsies. However, professional pathologists must expend considerable time and effort to accomplish this task. The process of breast cancer diagnosis is not only time-consuming and
∗ Corresponding author. E-mail address: [email protected] (S.-W. Lee).
Fig. 1. Examples of microscopic biopsy images.
expensive but also strongly depends on the prior knowledge of the pathologist and the consistency of pathologic reports.
The average diagnostic accuracy of pathologists is approximately 75% .
Fortunately, the development of computer vision and machine learning potentially offers more reliable classification methods for the histological assessment of hematoxylin and eosin stained sections. These methods can automatically clas-sify breast tissues into different categories with high classification rates. Thus, many researchers have developed fast and precise image analysis algorithms for breast cancer detection tasks. However, their results are still far from meeting ac-cepted clinical requirements. For this reason, researchers have been expending most of their efforts into the development of new algorithms for histopathological image analyses [18,26,41]. These algorithms aim to achieve the precise classification of breast tissues as normal tissues, nonmalignant (benign) tissues, in situ carcinomas, or invasive carcinomas. In the category of benign lesions, images show changes in the normal structures of breast parenchyma that do not progress to malignancy. In situ carcinoma indicates cells that are restrained inside the mammary ductal-lobular system. Unlike in situ carcinomas, invasive carcinomas present a profile where cells spread beyond the structure of the mammary ductal-lobular system.
One of the challenges in analyzing breast cancer histopathology is to deal with a wide variety of hematoxylin and eosin stained sections, which is attributed to differences among people, different protocols used in labs, skills of pathologists in scanning images, and different staining procedures . Fig. 1 shows some breast cancer histopathology images that are considered challenging for classification. Each of them belongs to one of four tissue classification categories, including normal tissues, nonmalignant tumors, in-situ carcinomas, and invasive carcinomas. To overcome this challenging problem, we developed an automated breast cancer detection method to classify breast tissues precisely into the four listed categories above. In particular, our proposed method is synthesized from several novel concepts.
In this research, we present an ensemble of deep convolutional neural networks (DCNNs) trained to extract the most useful visual features from multi-scale training images. The use of DCNNs increases the accuracy of classifying multi-label breast cancers by aggregating multi-scale contextual information. In addition, this network can extract both global and local information from original images. Conversely, by using lower-resolution input images, the receptive field of the network in the original image can be expanded to cover global features more adequately. Furthermore, higher resolution input images are used in our network to extract multi-scale features in local regions. These advantages can be applied effectively in breast cancer detection tasks because breast cancer tumors and cells have a wide variety of shapes, sizes, margins, and densities. First, our network can detect differently sized breast cancer tumors by extracting multi-scale local features that are very important for doctors to determine the stage of breast cancer. In most cases, if the doctor only detects small tumors, the patient has better chances for long-term survival. Otherwise, a larger cancer can be more aggressive and doctors may recommend a mastectomy or chemotherapy before surgery. Second, our network can recognize characteristic abnormalities in breast cancer tissues, such as shapes and margins, based on the detailed information obtained in local regions. The shapes of breast cancer tumors are usually round, oval, lobular, or irregular. Poorly defined or spiculated margins are often worrying signs of breast cancer cells. Most breast lesions and tumors have ill-defined borders and certainly need more investigations. Finally, this ensemble of deep convolutional neural networks is able to extract global information that is used to estimate the tumor densities and the number of breast tissues. This is an important advantage because a high-tumor density in terms of the amounts of fatty elements usually constitutes a highly suspicious sign for breast cancer.