The female reproductive system relies on the ovaries to produce eggs, but ovarian cysts can lead to complications such as torsion, infertility, and cancer, making it essential to diagnose them quickly. Ultrasound images are commonly used to detect ovarian cysts, but segmenting cyst regions from the surrounding tissue poses a challenge due to complex patterns and similar intensities. Few methods use the background's texture information to facilitate foreground segmentation. Ultrasound images include characters like speckle noise, low contrast appearance, and blurry boundaries that further complicate the task. Lesion shape and position variations exacerbate these challenges. This study proposes an improved deep learning-based segmentation technique using a database of ovarian ultrasound cyst images to overcome these issues. At the outset, the input has undergone pre-processing using non-sub-sampled contourlet domain-based cross-guided bilateral filtering (CGBF) and improved U-Net (IU-NET) for image segmentation. The presented architecture involved reducing the intricacy of U-Net through the alleviation of certain parameters. This resulted in a substantial acceleration of the learning process, by a factor of 100. To optimize the improved U-Net model, the Seagull Optimization Algorithm (SOA) was used. The algorithm helped to fine-tune the hyper-parameters of the U-Net architecture, including the batch size, learning rate, and epoch count, to achieve optimal performance. The optimization was performed by solving an objective function, which involved determining the dice loss coefficient (DLC) and weight cross-entropy (WCE). A numerical analysis was conducted, which demonstrated that the proposed methodology outperforms existing methods in terms of segmentation accuracy. The proposed model achieved a pixel accuracy of 99.36%, which was significantly higher than that achieved by existing methods.