Comparative Analysis of Different U-Net Variants for Breast Cancer Detection in Histological Images

Abstract

One of the most common forms of cancer is breast cancer, and the key to treating it is early detection. First and the most important phases of treating breast cancer is an accurate diagnosis. Numerous studies on predicting the type of breast cancers may be found in the literature. This study used information on breast cancer tumors from the BreCAHAD dataset to predict the types of breast tumors. To discover the region of interest, we used two distinct CNN architectures: U-Net and its variant U-Net++. With 93.15% accuracy, U-Net++ model shows promise for breast image segmentation. Shape and Tetrolet characteristics are used for feature extraction. This study examines how well VGGNet and VGGNet adjusted using Remora Kill Herd Optimization (RKHO) architectures detect breast cancer by analyzing histopathology images. We create reliable and accurate deep learning models by carefully adjusting optimizers and hyperparameters. According to our research, the VGGNet model trained using the RKHO performs exceptionally well in terms of metrics like accuracy, sensitivity and specificity. Corresponding scores obtained are 0.925, 0.915, and 0.929.