Deep convolutional neural network-based enhancement of Digital Breast Tomosynthesis images

Digital Breast Tomosynthesis (DBT) has emerged as a powerful imaging modality for early breast cancer detection, particularly in women with dense breast tissue. DBT captures multiple projection images over a limited angular range, enabling a quasi-3D reconstruction of the breast. However, this limited-angle acquisition introduces image artifacts, most notably blurring and low contrast in the reconstructed slices, which can obscure critical diagnostic features such as microcalcifications, which are early indicators of breast cancer. As a result, diagnostic accuracy is compromised, and the workload on radiologists increases significantly due to the need to manually examine a large number of slices. To address these challenges, this research presents a novel deep learning-based image enhancement framework designed to enhance the visibility of microcalcifications in DBT images. The study is divided into two main phases. In the first phase, a hybrid blur detection model, referred to as the Convolutional Neural Network–Support Vector Machine–Blur Factor (CNNSVM-BF), is developed by combining features extracted from a lightweight Convolutional Neural Network (CNN) with a handcrafted Laplacian-based Blur Detection (LbBD) algorithm. These hybrid features are classified using a Support Vector Machine (SVM), which enables the accurate detection of blurry slices and allows for the exclusion of low-quality images from subsequent analysis. This selective approach not only reduces computational burden but also mirrors the practical workflow of radiologists who often ignore visibly blurry slices. In the second phase, non-blurry (sharp) slices are subjected to a microcalcification contrast enhancement pipeline that includes two steps; first, an Unsharp Masking Very Deep Super Resolution (UMVDSR) model is applied to improve global image clarity. Second, a newly proposed Microcalcification Contrast Enhancement (McCE) algorithm is introduced as a post-processing step, specifically designed to enhance the contrast and visibility of microcalcification regions, enabling better lesion detectability while preserving anatomical structures. The proposed Deep CNN-based enhancement framework is evaluated using publicly available Breast Cancer Screening Digital Breast Tomosynthesis (BCS-DBT) dataset, as well as a custom dataset collected at the Advanced Medical and Dental Institute Digital Breast Tomosynthesis (AMDI-DBT). The results demonstrate substantial improvements in both microcalcification visibility and overall image quality, achieving a peak signal-to-noise ratio (PSNR) of 47.6454 and a structural similarity index measure (SSIM) of 0.9995. Compared to baseline methods, the proposed approach exhibits a significant improvement in image clarity and microcalcification detectability. These quantitative results, supported by qualitative assessments from expert radiologists, confirm the statistical significance of the enhancement and underscore the potential integration of the proposed intelligent system into clinical CAD workflows for more accurate and early breast cancer diagnosis. In conclusion, the proposed Deep CNN-based enhancement framework holds promise for integration into computer-aided diagnosis systems, supporting early detection of breast cancer in clinical practice.