Medical image segmentation plays a crucial role in facilitating clinical diagnosis and treatment, yet it poses numerous challenges due to variations in object appearances and sizes with indistinct boundaries. This paper introduces the MHSAttResDU-Net architecture, a novel approach to automatic medical image segmentation. Drawing inspiration from the double U-Net, multi-head self-attention (MHSA) model, and residual connections, the proposed model is trained on images pre-processed by the innovative ranking-based color constancy approach (RCC). The MHSAttResDU-Net includes the integration of RCC to control model complexity and enhance generalization across diverse lighting conditions. Additionally, the incorporation of the sparse salient region pooling (SSRP) unit in the encoder-decoder blocks reduces the dimension of feature maps, capturing essential local and global channel descriptors without introducing learnable parameters. MHSA gates are strategically employed in both down-sampling and up-sampling paths, allowing the recollection of additional relevant dimensional data. This effectively addresses dissimilar feature representations, minimizing unfocused noise and artifacts while reducing computational costs. Furthermore, Leaky ReLU-based residual connections between the encoder and decoder enhance the model’s capability to recognize complex shapes and structures, ensuring improved gradient flow and faster convergence. Experimental results demonstrate the superiority of the MHSAttResDU-Net architecture across diverse datasets, including COVID-19, ISIC 2018, CVC-ClinicDB, and the 2018 Data Science Bowl. The model achieves state-of-the-art performance metrics, including an accuracy of 99%, representing a promising advancement in automated medical image analysis with potential implications for improving patient care and diagnostic accuracy.
