Medical image segmentation is a branch of computer vision that is essential to diagnose diseases and improve patient care. In this study is focused on the dual-channel UNet (DC-UNet) model to detect polyps. Polyps are lesions that vary in size and are key to preventing colorectal cancer through early identification. To train and evaluate the DC-UNet, we use the challenging CVC-ClinicDB dataset, which includes colonoscopy video frames with annotated polyp regions. Given the complexity of this dataset, we propose using genetic algorithms to optimize the DC-UNet model’s hyperparameters. Specifically, we tune the selection of gradient-based optimizers, dropout rates, learning rates, \(\beta _1\) for first-order moment, and \(\beta _2\) for second-order moment. We tune these model hyperparameters to achieve optimal results, improving its performance by 4. 36% compared to the standard hyperparameter values. This approach offers a promising direction to improve medical image analysis and outcomes in early cancer detection.