Histopathological breast cancer classification using CNN

Abstract

In the year 2021, there were 2.26 million new diagnoses of breast cancer cases among women globally. Over the years, various imaging methods such as mammography, Magnetic Resonance Imaging (MRI), Positron Emission Tomography (PET), X-ray, and Computer Tomography (CT) have been used as helpful diagnostic tools for cancer prognosis. However, these methods expose the patient to rays that have the potential to be cancerous. Added to the cancerous risk, due to the fatigue experienced by histopathologists, many errors can be introduced in such diagnosis. Recently, Computer Aided Diagnosis (CAD) via Convolutional Neural Network (CNN) has been employed to overcome the fatigue and decrease in concentration experienced by histopathologists and clinicians during a cancer diagnosis. Though CNN has shown better results than conventional cancer diagnosis methods, using a single dataset for CNN training has been an accuracy challenge in diagnosing breast cancer. This study proposed a classification of histopathological images of breast cancer using CNN. A hybrid dataset was curated by preprocessing the histopathological dataset from two popular datasets, namely, BreakHis and Histo to the best of our knowledge this combination of is new. The cancer tissue images were classified as either Malignant or Benign. Instead of using a model with random weights from scratch, this study uses transfer learning to build four pre-trained models, namely DenseNet201, ResNet50, ResNet101 and MobileNet-v2. Five protocol setups for the dataset were set up and divided into 70% and 30% for training and testing, respectively. Of the four pre-trained models, DensNet201 achieved an accuracy of 91.37% and a sensitivity level of 100% at a 200x magnification factor. Our best-performed model outperforms state-of-the-art models by achieving an accuracy of 91.37%.