A Deep Learning Based Breast Cancer Classification System Using Mammograms

An automatic breast cancer detection and classification system plays an essential role in medical imaging applications. But accurate disease identification is one of the complicated processes due to the existence of noisy contents and irrelevant structure of the original images. In conventional works, various medical image processing techniques have been developed for accurately classifying the types of breast cancer. Still, it confronts difficulties due to the aspects of increased complexity in computations, error values, false positives, and misclassification outputs. Hence, this research work proposes to develop an optimization-based classification system for the breast cancer identification system. Here, the Gaussian filtering and Adaptive Histogram Equalization (AHE) techniques are utilized for preprocessing the original mammogram images by eliminating the noisy contents and enhancing the contrast of an image. Then, the Markov Random Adaptive Segmentation (MRAS) technique is employed for detecting the boundary region based on the random value selection. To make the classifying procedure easier, the set of features is optimally extracted from the segmented region with the help of a Genetic Algorithm (GA). In which, the global best fitness value is estimated by using the crossover, mutation, and selection operations. Finally, the Convolutional Neural Network (CNN) classification technique is utilized for categorizing the image as to whether normal or abnormal with its type. The entire performance analysis of the suggested model is validated and compared using multiple measures during the evaluation. In the proposed method GA performs feature selection and prunes unnecessary features. The major goal is to improve the classification performance while reducing the number of features used. The proposed system GA-CNN provides improved performance results with a reduced error rate.The suggested GA-CNN increases accuracy (98.5), sensitivity (99.38), and specificity values (98.4) as compared to the existing technique by effectively identifying the classed label.