This study examines the limitations of SAM (Meta) in medical image segmentation and optimizes its prompts using genetic algorithms to reduce user intervention. In pulmonary CT scans, chest X-rays, and spleen MRIs, it significantly improves accuracy (Dice coefficients of 97.17%, 96.57%, and 95.07%) without retraining or modifying SAM, and approaches the state of the art. However, it exhibits a performance ceiling that further iterations do not exceed. Broader clinical validation is still needed.
