Comparative Analysis of Optimizers for Breast Cancer Segmentation in Ultrasound Images Using U-Net

This paper presents a comparative study on the impact of various optimizers Adam, SGD, AdaGrad, and RMSprop on U-Net-based segmentation of breast cancer in ultrasound images. Using a synthetic dataset of breast ultrasound (BUS) images, generated via Data Expansion Network (DEN) and Probability Map Generator (PMG), the performance of each optimizer on segmentation metrics such as Dice coefficient, IoU, and accuracy are evaluated. Results demonstrate that the Adam optimizer yields the highest segmentation accuracy, while other optimizers exhibit specific strengths that may benefit different aspects of training stability and convergence.