Classifying Breast Cancer Using Deep Convolutional Neural Network Method

Musfequa Rahman, Kaushik Deb, Kang-Hyun Jo

Abstract: The efficacy of conventional classification systems is contingent upon the accurate representation of data and a substantial portion of the effort invested in feature engineering, which is a laborious and timeconsuming process requiring expert domain knowledge. In contrast, deep learning has the capacity to automatically identify and extract discriminative information from data without the need for manual feature creation by a domain expert. In particular, Convolutional Neural Networks (CNNs), a type of deep feedforward network, have garnered attention from researchers. This study conducts several preliminary experiments to classify breast cancer histopathology images using deep learning, given the small number and high resolution of training samples. The proposed approach is evaluated on the publicly available BreaKHis dataset, utilizing both a scratch model and transfer learning pre trained models. A comparison of the proposed scratch method to alternative techniques was carried out using a suite of performance evaluation metrics. The results indicate that the scratch model, with its independent magnification factor, achieved greater accuracy, with a binary classification accuracy of 99.5% and a multiclass classification accuracy of 96.1%.