DeepMixNet : Deep Multi‐Scale Interactive Feature Mixing Network for Automated Skin Lesion Segmentation

Abstract

ABSTRACT In recent years, medical image segmentation has emerged as a pivotal technology in medical image analysis, playing a particularly critical role in the diagnosis of skin diseases, including melanoma. Consequently, enhancing the accuracy and robustness of medical image segmentation remains a core challenge in contemporary medical image analysis. U‐Net is known for its efficient image processing capabilities and outstanding medical image segmentation performance. Therefore, it has been widely adopted in medical image segmentation. However, many current modeling approaches remain imperfect and exhibit substantial limitations when applied to real‐world clinical scenarios. In this paper, we propose DeepMixNet, a novel deep multi‐scale interactive feature fusion network specifically tailored for automated skin lesion segmentation. Specifically, we introduce DMixblock, a deep multi‐scale interactive feature mixing block integrated into the U‐shaped model, enabling low‐level spatial details and high‐level semantic information to interactively enhance each other through bidirectional paths. We conducted comparative experiments on three public skin lesion datasets (ISIC 2017, ISIC 2018, and HAM10000) and then used the PH 2 dataset as external validation. The experimental results show that the proposed DeepMixNet model demonstrates significant accuracy advantages in skin lesion segmentation tasks.