An automated system described in Japan demonstrated it can detect pneumothorax in newborns with high accuracy, addressing a clinical blind spot where subtle imaging findings are easy to miss. The approach was built and evaluated as a deep learning model for neonatal chest imaging. Separate work proposed entropy-based methods to decide when surgical AI answers should be trusted, tackling the practical problem of overconfident model outputs when accuracy is uncertain. The study frames calibration as a safety layer for real-time clinical decision support. Together, the items move AI from performance-only metrics toward operational risk management—accuracy plus calibrated trust.