Image Pre-Processing Techniques to Improve DBT Classification Accuracy With CNN ResNet-50

Abstract

Digital Breast Tomosynthesis (DBT) is a three-dimensional imaging technology that is meant to determine breast cancer more accurately than regular mammograms. But low contrast and noise can make DBT images less clear, which can make deep learning-based classification less accurate. This study seeks to assess the impact of diverse preprocessing techniques on breast cancer classification outcomes utilizing the ResNet-50 Convolutional Neural Network (CNN) model. We looked at four different types of preprocessing: The original image, CLAHE (contrast limited adaptive histogram equalization), color mapping, and Region of Interest (ROI). All DBT images were transformed into 2D format, and oversampling was conducted to rectify the disparity in the quantity of benign and malignant classes. We trained each image variant on the same ResNet-50 architecture and then checked its accuracy, precision, recall, and F1 score. The color mapping technique produced the most accurate results in experiments, with an accuracy of 95% and a consistent and balanced F1 score across all classes. These results show that adequate preprocessing has a big effect on how well CNN models work for DBT image classification and may also make automated breast cancer early detection systems stronger. Supplementary experiments combining ROI segmentation with CLAHE and with color mapping showed that this benefit is combination-dependent: ROI + CLAHE raised accuracy from 86 to 92%, whereas ROI + color mapping matched color mapping alone in overall accuracy (95%) but shifted the balance between classes, improving benign-class metrics at the cost of a small reduction in malignant-class recall indicating that combining techniques helps most when the underlying technique is weaker rather than uniformly improving an already strong one.