<scp>DenseUNet</scp> Architecture With Asymmetric Kernels for Automatic Segmentation of Medical Imaging

ABSTRACT Medical imaging is a core component of modern healthcare, essential for early disease diagnosis and effective treatment planning. Deep learning has emerged as a powerful tool to enhance medical image analysis, particularly in tasks such as segmentation, which is essential for identifying and delineating anatomical structures. A notable segmentation challenge involves accurately detecting narrow and elongated features. A novel DenseUNet architecture, enhanced with asymmetric convolutional kernels and a squeeze‐and‐excitation block, is proposed. It is specifically designed to adapt to such shape characteristics. The iterated local search metaheuristic is employed to optimize the kernel size within a search space of , and a squeeze‐and‐excitation block is integrated to enhance feature recalibration and network efficiency. The best‐performing asymmetric kernel achieved a processing time 5423 s faster than that of the conventional kernels. The proposed architecture is evaluated using the Dice coefficient and benchmarked against state‐of‐the‐art architectures using three databases (TMJ: Temporomandibular joints, DCA1: Coronary arteries, and ICA: Coronary arteries), achieving Dice scores of 0.7800, 0.8231, and 0.8862, respectively. These enhancements demonstrate improved segmentation performance and contribute to the development of more accurate and robust medical imaging tools.