Classification of Breast Histopathology Images using Color Deconvolution and DenseNet Architecture

Abstract—Several mammography and other imaging
techniques are widely used for breast cancer assessment, but
accurate detection is achieved by analyzing breast tissue
samples in the form of histopathological images. Histopathology image-based breast cancer assessment is highly challenging due to the complexities involved in the processing of images. Though deep learning methods are increasingly used in classifying breast cancer histopathological images, they face many issues
such as overfitting, high computational complexity and limited feature extraction capability. This work aims to classify breast histopathology images by applying color deconvolution based preprocessing combined with pre-trained deep learning models. Color deconvolution is applied on histopathological images to effectively handle the stain components associated with them. Furthermore, few deep learning models are deployed for classification and the best deep learning model is chosen for improving classification accuracy and other performance metrics. Experimental results reveal that the proposed DenseNet deep learning framework combined with color deconvolution performs better than several other transfer learning models in terms of detection accuracy, generalization ability, and other performance metrics.