Comparison of Projection and Reconstruction Techniques in Sinograms for Breast Lesion Classification

Abstract

Breast cancer remains one of the leading causes of mortality among women worldwide, and early detection and accurate staging of breast lesions are critical for effective treatment. Digital Breast Tomosynthesis (DBT) has emerged as a promising imaging technique, offering clearer, more layered breast images than conventional mammography. DBT generates sinograms from thin-layer projections, which are then used to reconstruct three-dimensional images. However, the reconstruction process can introduce artifacts, potentially leading to information loss and inaccuracies in lesion detection. This study compares the efficacy of direct analysis of preprocessed sinograms versus reconstructed images for breast lesion detection using Convolutional Neural Networks (CNNs). Specifically, we evaluated sinograms from 180-degree projections versus those from 360-degree projections and reconstructed images using simple back projection. The results demonstrate that 180-degree sinograms, when preprocessed for contrast enhancement, significantly outperform 360-degree sinograms and reconstructed images in terms of accuracy, recall, and F1 score. The superior performance of 180-degree sinograms underscores their potential as a viable alternative to traditional image reconstruction methods, offering a more effective approach to lesion detection and classification. This study contributes to advancing breast cancer diagnosis by highlighting the advantages of using preprocessed sinograms. It suggests further exploration of advanced image processing techniques and neural network architectures to improve diagnostic accuracy.