DB-DANet: A Dual-Branch Dense Attention Network with Charbonnier Loss for Low-Dose CT Image Reconstruction

Advancements in computer vision have significantly impacted medical image processing, particularly noise reduction in low-dose computed tomography (CT) scans. Recent progress in this field has seen deep learning techniques play a pivotal role in addressing this challenge, often relying on powerful network architectures and meticulously designed loss functions. In this paper, we propose DB-DANet, a novel dual-branch neural architecture designed to capture complementary features, enhancing model learning and reconstruction abilities. We introduce adaptive dynamic convolution block (ADCB) in the local information extraction layer to preserve fine details and texture features of the image. In the global information extraction layer, we propose dense multi-scale edge enhancement attention block (DMEAB), which perform feature fusion through dilated convolution and multi-dimensional attention mechanisms. Additionally, we propose integrated multi-scale attention block (IMAB) to enhance feature integration and robustness by refining both spatial and channel-wise feature selection, allowing the model to prioritize diagnostically significant features. The model was evaluated using the AAPM-Mayo Clinic LDCT Grand Challenge Dataset and demonstrated superior performance compared to previous state-of-the-art (SOTA) models across various architectures, achieving a PSNR of 33.4924, SSIM of 0.9159, and RMSE of 8.6469, all while maintaining a low number of learning parameters 1.33M. Our approach proves competitive with advanced LDCT denoising techniques, both in terms of visual quality and quantitative performance.