QED-UNet: Efficient Channel Attention and Decaying Warm Restart Cosine Annealing-Enhanced Quantum-Assisted UNet

Medical image segmentation is crucial for computer-aided diagnosis, providing pixel-level delineation of lesions and directly impacting subsequent quantitative analysis and clinical decisions. With the rapid progress of quantum computing, an increasing number of studies have integrated it with deep learning for feature extraction, thereby improving segmentation efficiency and accuracy. However, conventional U-Net-based models still suffer from limited feature representation in complex segmentation scenarios and existing quantum-assisted U-Net (QU-Net) designs mainly focus on quantum feature mapping while overlooking channel discrimination and training stability. We thus propose a hybrid quantum–classical-enhanced convolutional neural network, named QED-UNet, which integrates quanvolutional preprocessing, an efficient channel attention module, a decaying warm restart cosine annealing learning rate schedule, and a classical U-Net backbone. Specifically, the quanvolutional preprocessing is a fixed, non-trainable front end that maps local image patches into a nonlinear feature space. The ECA module improves channel-wise feature discrimination, enabling the model to focus more on lesion-related information while suppressing background noise and irrelevant features. The DWRCA strategy dynamically adjusts the learning rate during training, improving convergence stability and optimization efficiency. We evaluate the performance of QED-UNet on ISIC2017 and ISIC2018 datasets. On ISIC2018, QED-UNet achieves Dice and IoU scores of 84.73% and 74.72%, exceeding the original QU-Net by 1.77 and 1.37 percentage points, respectively. These results indicate that combining quantum feature extraction, channel-wise feature refinement, and stable learning rate scheduling can boost hybrid quantum–classical medical image segmentation performance.