MH-UNet: A Modified Hybrid UNet with Attention and Graph Neural Networks for Leukemia Cell Segmentation
International journal of intelligent engineering and systems · 2025
Medical image segmentation plays a crucial role in disease diagnosis and treatment planning by enabling precise identification of anatomical structures and abnormalities.In haematology, leukaemia cell segmentation is essential for distinguishing malignant cells from normal blood cells in microscopic images, aiding in early diagnosis and prognosis.However, traditional segmentation methods struggle with challenges such as poor boundary delineation, intra-class variations, and the presence of overlapping cells.Deep learning-based segmentation models, particularly UNet and its variants, have significantly improved biomedical image analysis.Despite their success, conventional UNet architectures face limitations such as vanishing gradients in deep networks, insufficient global feature extraction, and difficulty in handling complex cellular structures.To address these challenges, we propose a Modified Hybrid UNet that integrates advanced deep learning techniques to enhance segmentation accuracy and robustness.Our model incorporates Residual Blocks (Res-UNet) to improve gradient flow and feature propagation, Attention Mechanisms (Attention UNet) to focus on leukemia cells while suppressing background noise, and Efficient Channel Attention (ECA) to dynamically refine feature representations.Additionally, to enhance global context understanding, we introduce a Transformer-based Global Context Module, improving long-range dependencies, and Graph Neural Networks (GNNs) to ensure structural consistency and better object segmentation.By leveraging these advancements in deep learning model, the proposed model effectively balances local and global feature learning, improving segmentation precision and computational efficiency.Experimental evaluations demonstrate that the proposed has reported the segmentation validation accuracy of 99.20% by using ALL-IDB dataset.