A Comprehensive Study on Medical Image Segmentation using Deep Neural Networks
This study addresses three critical challenges in medical image segmentation (MIS): weak interpretability of deep neural network (DNN) models, fragmented evaluation frameworks, and insufficient clinical trustworthiness. To tackle these, we systematically introduce the DIKIW (Data–Information–Knowledge–Intelligence–Wisdom) hierarchy into MIS evaluation for the first time. We propose an eXplainable AI (XAI)-driven “Intelligence-to-Wisdom” evolution pathway, integrating multi-level semantic modeling with clinically grounded assessment to establish a comprehensive DIKIW-aligned method taxonomy. Key bottlenecks—including black-box decision-making and inadequate early-lesion representation—are identified, and deployable transparency-enhancing solutions are provided. Our framework significantly improves both interpretability and early detection rates for cancerous lesions, offering theoretical foundations and practical guidelines for high-assurance clinical AI systems.