Breast ultrasound image segmentation to detect breast cancer early is important but difficult due to speckle noise and anatomy variations. Though deep learning models have enhanced segmentation accuracy, the impact of optimizer selection has been comparatively less researched. This work proposes a training-centered strategy grounded on the Marine Predators Algorithm (MPA) to fine-tune the Lookahead optimizer by selecting the best base optimizer and tuning crucial hyper- parameters (learning rate, synchronization period α, and interpolation factor k). Experimental results on the BUSI dataset with U-Net and Ef- ficientNetB7 encoder show that MPA-optimized configurations are much better than usual. The optimum three settings (Opt4, Opt3, and Opt6) were at Dice measures of 78.78%, 78.64%, and 78.60% respectively, categorically affirming that metaheuristic tuning is optimum for improving segmentation accuracy.