Accurate segmentation of ovarian tumors in ultrasound images is essential for early diagnosis and treatment planning. However, this task remains challenging due to the inherent limitations of ultrasound imaging, such as noise, shadowing, low contrast, and blurred tumor boundaries. Furthermore, the large variability in tumor shape, size, and position across patients further complicates traditional segmentation methods. Most existing deep learning approaches rely solely on visual features, overlooking the potential benefits of incorporating semantic information. In this study, we propose a novel multimodal segmentation framework that explicitly integrates both visual and textual information. Our architecture is based on a UNet-style backbone enhanced with a pretrained VGG16 encoder and a Spatial Pyramid Pooling Fast (SPPF) block for improved multi-scale feature learning. The key contribution lies in the integration of Image-Text Matching (ITM) modules into the segmentation pipeline, enabling deep cross-modal interaction between image features and semantic tumor descriptions. Additionally, we introduce a lightweight captioning module that automatically generates textual descriptions from ultrasound images, allowing the model to be trained even when manual annotations are unavailable. To further enhance performance, we employ a combination of loss functions including Tversky, Weighted Binary Cross Entropy (WBCE), and Structure Similarity (SSIM), aiming to balance precision, recall, and structural similarity. Experimental results on the OTU_2d dataset demonstrate the effectiveness of our approach, achieving 90.16% Dice, 82.05% IoU, 90.23% Recall, 91.12% Precision.