Anatomically ROI- and IOI-informed Hybrid U-Net model for Abdominal Object Segmentation in CT images

Segmentation in medical imaging is a critical and challenging task, essential for improved disease diagnosis, staging, treatment planning, and response assessment. Over the years, various deep learning (DL) methods have been developed to enhance image segmentation. Among these, encoder-decoder methods, particularly the U-Net model and its variants, have proven effective in medical image segmentation. These methods incrementally address the limitations of artificial intelligence (AI) in recognizing high-level object concepts by integrating different-level features. We add natural intelligence (NI) into AI and leverage anatomical knowledge to further improve object segmentation accuracy. Our study introduces the hybrid U-Net (HUN) model that focuses on regions of interest (ROI) and intensities of interest (IOI) determined by NI, thus enhancing resource utilization efficiency and reducing irrelevant information. The HUN model balances local and global feature extraction, improving segmentation outcomes in 15 of 16 objects of interest in abdominal low-dose CT images compared to the U-Net. The novel concept of hybrid intelligence (HI), which combines NI with AI/DL techniques, is explored, showcasing its robustness and generalizability across various medical imaging applications. Our findings underscore the importance of incorporating NI into AI-driven medical imaging methods, paving the way for more effective and reliable solutions in medical image analysis and holding significant potential for improving patient care and clinical outcomes.