Early detection of lung cancer is critical, but accurate diagnosis via chest X-rays is challenging due to variability in physician experience and fatigue. Medical AI systems have potential to assist diagnosis but suffer from limited trust because they lack comprehensive clinical knowledge, leading to diagnostic uncertainty. Unlike physicians who use extensive background knowledge, AI models typically rely solely on repetitive learning of lesion data, which restricts their diagnostic confidence. To address this, the study introduces an Uncertainty-Aware Learning Policy that incorporates physicians’ background knowledge alongside lesion information during training. Using a dataset of 2,517 lesion-free and 656 nodule images from Ajou University Hospital, the proposed model improved sensitivity by 10% compared to a baseline while reducing uncertainty as measured by entropy. This advancement suggests that integrating clinical knowledge into AI can enhance diagnostic reliability and potentially increase physician trust. The reduction in uncertainty is crucial for clinical adoption, as it addresses skepticism about AI’s diagnostic confidence. The findings imply that future AI diagnostic tools should consider knowledge-aware learning to improve performance and acceptance. Further validation and refinement could lead to widespread clinical use, improving early lung cancer detection and patient outcomes.
👉 Pročitaj original: arXiv AI Papers
