A Study and Analysis of Deep Neural Networks for Cancer Using Histopathology Images

A Study and Analysis of Deep Neural Networks for Cancer Using Histopathology Images

Anu Singha, Jayanthi Ganapathy
DOI: 10.4018/978-1-6684-4405-4.ch002
(Individual Chapters)
No Current Special Offers


Due to the complexity of histopathology tissues, an accurate classification and segmentation of cancer diagnosis is a challenging task in computer vision. The nuclei segmentation of microscopic images is a key prerequisite for cancerous pathological image analysis. However, an accurate nuclei segmentation is a long running major challenge due to the enormous color variability of staining, nuclei shapes, sizes, and clustering of overlapping cells. To address these challenges and early diagnosis as well as reduce the bias decisions of expert lab technician of cancer in clinical practice, the authors study the classification of computer-aided frameworks and automatic nuclei segmentation frameworks based on histopathology images by convolutional deep learning. The authors have used a publicly available PatchCamelyon and 2018 Data Science Bowl histology image dataset for this study. The results are compared and expected to be useful clinically for technician experts in the analysis of cancer diagnosis and the survival chances of patients.
Chapter Preview


In health concerns, the cancer diseases become the most common and life-threatening issue. In India, the incidence of any cancer growth rate is increases in young age group at very aggressive Shah, 2020). Cancer starts from a benign state and at the early stages without proper treatment, it becomes malignant when the cells start to grow abnormally. It is started to form a mass or lump via divide infected cells more frequently than normal healthy cells. However, if exposed early, the cancer is a greatly curable disease with 97% chances of survival (Ng, 2009).

The availability of appropriate screening devices is essential for detecting the initial symptoms of cancer. Numerous imaging techniques are used for the screening to detect this disease such as histopathology. To improve the accuracy of the diagnosis for patients, the histopathology images are considered as the gold standard among other imaging techniques (Ng, 2009). Moreover, the histopathological examination can deliver more inclusive and reliable evidence to diagnose cancer and measure its effects on the surrounding tissues (Gurcan et al., 2009; Hipp et al., 2011; Pickles et al., 2015). In this chapter, we have examined these histopathology images.

For analysis of histopathology image, the detection and segmentation of nuclei cells are essential steps. These segmented nuclei are used in the grading diagnosis of many cancers which require comprehensive analysis of the characteristics of the nuclei such as shape, size, gray value, color variation of samples, clusters of nuclei with overlapping, and ratio of nuclei to cytoplasm. It is a major challenge in histopathology images of different patients where the shape and appearance of the different nuclei for disease stages vary greatly. In breast cancer, the identification of the stages of aggressiveness of the disease based upon the Nottingham Histologic Score system which also largely based off the morphologic attributes of the histopathology nuclei (Basavanhally et al., 2011). As a consequence, the accurate segmentation of histopathology image nuclei is a great challenging work in developing automated machine via computer assisted decision support for digital histopathology. Other than these, the segmentation of nuclei of histopathology image is essential to numerous studies, such as feature extraction, cell counts, and classification.

In this chapter, we have examined cancer histopathology images in two ways. First, we examine the patch-based histopathology images for classification either carcinomas metastasis/malignant patch or non-carcinomas/normal patch. Secondly, deep convolutional-de-convolutional segmentation approaches will be analysed to handle accurate segmentation of nuclei.

The contribution of this study will as follow:

  • 1.

    We have studied the benchmark CNN models for abnormality classification. Unlike the conventional method of using texture feature-based analysis, the benchmark models make use of the metastatic cancer in small image patches taken from larger digital pathology scans to signify the abnormality.

  • 2.

    The classification performance of the abnormality classification system has been evaluated with a public PatchCamelyon (PCam) (Bejnordi et al., 2017) histopathologic cancer database by using a series of state-of-the-art classification benchmark systems. The performances will be measured through supervised evaluation metrics like accuracy, recall, precision, F1 score, confusion matrix, and ROC AUC.

  • 3.

    Encoder-Decoder models has studied to handle nucleus segmentation for breast cancer histopathology images. In Encoder part, it stitches feature maps in the channel dimension to achieve feature fusion and uses a skip structure in Decoder part to combine low- and high-level features to ensure the segmentation effect of the nucleus.

  • 4.

    The performance of nuclei segmentation has been done with a publicly available dataset i.e. 2018 Data Science Bowl (Kaggle, n.d.).

Key Terms in this Chapter

Nuclei: The nuclei of a cancer tissues are larger and darker than that of a normal tissue. Another feature of the nucleus of a cancer tissue is that after being stained with certain dyes, it appears darker when seen under an optical microscope.

Histopathology: Histopathology is a microscopic examination of a biopsy specimen that is processed onto glass slides. This examination diagnoses the signs of the cancer disease.

Invasive Cancer: The invasive breast cancers may have spread within the breast only, or to nearby lymph nodes. They may have spread to distant body parts.

Data Augmentation: These are techniques which is used to increase the amount of image/data by adding slightly transformed/modified copies of already existing images/data or newly created synthetic images/data from existing images/data.

Metastatic Cancer: The metastatic breast cancers have spread outside of the breast and nearby lymph nodes to distant body parts.

Classification CNN Models: Classification CNN models that classify each pixel to identify what is in an image. For example, VGG, ResNet, and AlexNet.

Segmentation CNN Models: Segmentation CNN models provide the exact outline of the region of objects, i.e., pixel by pixel details are provided for given objects in an image. For example, U-Net and PSPNet.

Complete Chapter List

Search this Book: