Automatic Diagnosis of Breast Cancer from Histopathological Images Using Deep Learning Technique

Abstract

Breast cancer is an invasive tumor that develops in the breast tissue. It is the most common cancer and also the leading cause of cancer mortality among females worldwide. Survival from breast cancer can be increased through advances in screening methods and early diagnosis. Clinical examination, screening using imaging modalities and pathological assessment (biopsy test) are common methods of breast cancer diagnosis. Among these, pathological assessment can be taken as a gold standard due to its potential in identifying the cancer type, grade and stage. However, current diagnosis using biopsy test is commonly done through visual inspection. This manual diagnosis is time consuming, tedious and subjective which may lead to misdiagnosis. Different machine learning and deep learning methods have been proposed in the literature to automate the manual breast cancer diagnosis mechanism. To our knowledge, an integrated system that can classify breast cancer to its subtypes and identification of the cancer grade from biopsy tests has not been previously done. Identification of breast subtypes and grade is essential for understanding the biological characteristics and clinical behavior as well as for developing personalized treatments. Moreover, the results claimed from previously proposed automation techniques are less accurate and unreliable. In this thesis, an automatic detection of breast cancer type, subtype and grade was proposed based on deep learning neural network model. The system was developed using python software. It was trained and validated using histopathological images acquired from „break-his dataset‟, the „zendo online dataset‟ and local data acquired from Jimma University Medical Center (JUMC), by using optikam PRO5 digital camera and optika microscope. All images have been preprocessed using CLAHE algorithm and histogram matching techniques, prior to feeding to the ResNet 50 pre-trained model. The developed system is capable of classifying breast cancer into binary classes (benign and malignant) and multi-classes (sub-types). Identification of cancer grade is done for cancers that are classified as malignant ductal carcinomas. Our test results showed that, the proposed method is 96.75%, 96.7%, 95.78%, and 93.86 % accurate for binary classification, benign sub-type classification, malignant sub-type classification, and grade identification, respectively.