<p>Automatic segmentation of ovarian structures and tumors in ultrasound imaging is essential to support early diagnosis and effective treatment planning. Despite significant progress in recent years, the problem remains challenging due to the strong variability and unclear boundaries of ovarian structures and tumors. This study proposes a segmentation framework, namely CSA-SegNet (Combined Spatial Pooling with Attention for Ovarian Tumor Segmentation Network), specially designed to address the above challenges of ovarian ultrasound images. It incorporates a VGG19 as an effective encoder for learning tumor-relevant features and a Spatial Pyramid Pooling-Fast Block (SPPF), embedded within an Attention U-Net architecture for multi-scale representation. To deal with the unclear boundaries issue, we propose a new loss function that enhances conventional loss functions such as IoU loss, SSIM Loss, and Focal loss function with Active Contour Euler Elastical loss. CSA-SegNet was evaluated on two publicly available datasets: OTU2D, USOVA3D, and one self-collected dataset OvaTUS-V1. Experimental results show that the proposed model achieves competitive performance with Dice scores of 87.56% on OTU2D, 88.64% on OvaTUS-V1, and 78.29% on USOVA3D. It consistently outperforms compared baseline methods, with Dice score gains ranging from 2.15% to 27.06%, 2.15% to 16.71%, and 2.14% to 7.40% on OTU2D, OvaTUS-V1, and USOVA3D, respectively, demonstrating its effectiveness in enhancing segmentation accuracy and boundary delineation. The source code of CSA-SegNet is available at <a href="https://github.com/phamthiloanhd/CSA-SegNet">https://github.com/phamthiloanhd/CSA-SegNet</a></p>