Chapter 12 - Efficient transfer learning techniques for breast cancer histopathological image classification

Abstract

Breast cancer (BC) is one of the most prevalent cancers across the globe. The second most common malignancy among women is breast cancer after lung cancer. A woman being affected by breast cancer is 1 in 39 or roughly 2.5%. The frequency of breast cancer-related deaths has been steadily declining since 1989, and this is because of various screening methods, treatments, improved awareness, and advanced computer-aided diagnosis systems to identify breast cancer are assumed to be the cause of decrease in mortality rates. To lower the number of fatalities, there is a significant need to detect breast cancer at an early stage. Biopsy is a common medical procedure that is commonly performed to investigate the abnormalities in the tissue by extracting a sample of tissue and examining it under a microscope; the image obtained from the microscope is a histopathological image. For the early detection of breast cancer, histological image analysis is crucial. Using classification algorithms, we can predict whether the tissue is malignant or benign. There are many existing techniques in machine learning like SVM and K-means algorithm and in deep learning like CNN with VGG, SegNet, DenseNet, and GoogleNet. We use transfer learning techniques to train the efficient time-consumed transfer learning approach on ResNet50 architecture to evaluate the effectiveness and accuracy of categorization performance. It would take much longer to train a model from the starting phase using initial random weights than it would to train a network using the fine-tuned model based on the knowledge obtained from other pretrained models, transfer learning approach. The suggested technique uses transfer learning techniques on histopathological images from the publicly accessible BreakHis dataset to categorize benign or malignant BC, which is shown to be superior in comparison to other deep learning approaches in the form of BC detection accuracy with strong transferability and computational speed.