RESUMEN
BACKGROUND AND OBJECTIVE: Breast cancer is a frequently diagnosed cancer in women, contributing to significant mortality rates. Death rates are relatively higher in developing nations due to the shortage of early detection amenities and constraints on access to technical advances combating this disease. The only way to diagnose cancer with certainty is through biopsy performed by pathologists. Computer-aided diagnostic algorithms can assist pathologists in being more productive, objective and consistent in the diagnostic process. The focus of this work is to develop a reliable automated breast cancer diagnosis method which can operate in the prevailing clinical environment. METHODS: Nuclei overlap and complex structural organisation of the breast tissue in biopsy images make nuclei segmentation, feature extraction and classification challenging. In this work, a nucleus guided transfer learning (NucTraL) methodology is proposed as a simple and affordable breast tumor classification algorithm. The image feature is represented by fusion of local nuclei features that are extracted using convolutional neural network (CNN) models pretrained on the ImageNet database. The nucleus patch extraction strategy used in this work avoids fine segmentation of the nuclei boundary but provides features with good discriminative power for classification. Classification of the fused features into benign and malignant classes is performed using a support vector machine (SVM) classifier. A belief theory based classifier fusion (BCF) strategy is then employed to combine the outputs arising from the different CNN-SVM combinations to improve accuracy further. RESULTS: Evaluation of results is achieved by executing 100 random trials with 70%-30% train to test division on the publicly available BreaKHis dataset. The proposed framework achieved average accuracy of 96.91%, sensitivity of 97.24% and specificity of 96.18%. CONCLUSION: It is found that the proposed NucTraL+BCF framework outperforms several recent approaches and achieves results comparable to the state-of-the-art methods even without using high computational power. This qualitative framework based on transfer learning can contribute significantly for developing cost effective and low complexity CAD system for breast cancer diagnosis from histopathological images.