D²U-Net: a dual-path hybrid UNet architecture for precise medical image segmentation

Many:medical image segmentation tasks heavily depend on U-shaped encoder–decoder networks like as U-Net, which have shown remarkable success. Nevertheless, these designs may have difficulties in effectively capturing inter-dependencies within hierarchical layers, resulting in problematic segmentation of lesions with irregular borders and intricate morphologies. In order to tackle these difficulties, we present D²U-Net, an innovative dual-path Hybrid U-Net architecture specifically implemented to improve feature interactions and optimize the use of multi-scale information. In order to capture both global and local features through deeper hierarchical interactions, the design utilizes dual encoder–decoder paths. Moreover, we present a Contextual-Spectral Fusion Module (CSFM) specifically developed as a comprehensive information fusion and enhancement approach for encoder–decoder approaches. This module facilitates the smooth and extensive sharing of information across multiple stages and paths. The proposed model integrates fine-grained texture details with high-level semantic features to improve segmentation accuracy. The D²U-Net is further enhanced by integrating global context and local edge extraction in a hybrid fashion through two encoder–decoder paths. The precise integration of multi-scale features is ensured by lightweight operations and efficient merging, resulting in the accurate segmentation of complex medical images. We assess the performance of D²U-Net on four intricate medical image segmentation tasks: the segmentation of skin lesions datasets, including the ISIC2018 and PH2 datasets, polyps, and brain tumor datasets. The proposed approach is consistently superior to state-of-the-art approaches, as evidenced by experimental results. This approach provides greater accuracy and generalization across an extensive variety of datasets. Our method achieved an IoU of 90.5% and a Dice coefficient of 95.0% on the ISIC2018 dataset. On the PH2 dataset, it achieved an IoU of 84.0% and a Dice coefficient of 91.2%. On the polyp dataset, it achieved an average IoU of 96.15% and a Dice coefficient of 97.95%. On the brain tumor dataset, it achieved an IoU of 89.4% and a Dice coefficient of 94.3%. It is anticipated that this method could be used in clinical practice in the future owing to its consistent accuracy throughout a wide range of datasets covering various medical conditions. The code is available in the GitHub repository at https://github.com/nooriahmed/D2U-Net-Hybrid-U-Net.