A hierarchical pipeline for preprocessing and classification of breast histopathology images into benign and malignant

Abstract

Breast cancer is one of the leading causes of death among women across the world. Currently, healthcare informatics is primarily centered on classifying breast cancer images, given the life-threatening nature of this disease. However, variations between and within observers can occur, which may result in misdiagnosis during cancer detection. The proposed work contains a framework for enhancement, segmentation of nuclei followed by classification of breast histopathological images. Enhancement is performed using gaussian filter and decorrelation stretch methods. Segmentation of nuclei is accomplished by extracting YCbCr components, enhancing Y component using Contrast Limited Adaptive Histogram Equalization (CLAHE) and applying K-means clustering technique. K-Nearest Neighbor (K-NN) and Support Vector Machine (SVM) classifiers are employed for automatic classification of breast histopathological images based on texture, geometry and YCbCr features. The proposed method is tested on publicly available BreakHis dataset and yielded best enhancement results for gaussian filter with highest PSNR and lowest RMSE values. The classifiers were assessed using performance metrics including accuracy, specificity, sensitivity and F-score. Experimental results revealed that the SVM achieved better performance than the existing methods, attaining nearly 87.5% accuracy on the dataset.