Multi-scale morphology-aided deep medical image segmentation

Medical image segmentation serves as a critical tool for healthcare professionals, enabling the precise extraction of Regions of Interest (ROIs) from clinical images at the pixel level. The evolution of computer vision and machine learning algorithms has streamlined this labor-intensive segmentation process, which traditionally necessitates domain expertise. The intrinsic challenges posed by clinical images - such as irregular shapes, varying sizes, low contrast, and intricate details within specific areas, contribute to the intricacy of the task. In response to these complexities, we propose a solution involving three modules grounded in morphological operations: the Multi-scale Morphological Closing Module, the Multi-scale Morphological Opening Module, and the Multi-scale Morphological Gradient Module. In contrast to conventional morphological operations, our approach involves learning structuring elements through a training process, enabling effective adaptation to the irregular shapes of ROIs. To cater to the diverse range of ROI sizes in clinical images, we introduce the concept of dilation rates within the structural elements of morphological operations. Our proposal extends to MorphUNet, a lightweight framework for medical image segmentation. This architecture integrates proposed modules with UNet, presenting a tight-coupled synergy between deep neural networks and multi-scale morphological operations. Efficacy is substantiated across diverse medical imaging datasets, spanning modalities, conditions, and ROI proportions. Extensive experimentation validated through widely recognized segmentation metrics underscores our model's superiority compared to fifteen state-of-the-art segmentation methods and baseline models.