Ovarian cancer (OC) is the eighth major cause of death among gynecological diseases, and its early detection is important for reducing mortality. Computer-aided methods help to identify and analyze ovarian tumors at an early stage, especially when using 2D ultrasound (US) images, which are widely used and safe for the patients. However, many existing models still fail to achieve strong performance and lack scalability when applied to real clinical data. To overcome these issues, this paper presents a new lightweight model for ovarian tumor segmentation and classification using US images. The study uses the multi-modality ovarian tumor ultrasound (MMOTU) dataset, which contains 1469 2D US images and 170 contrast-enhanced ultrasonography (CEUS) images with both pixel-level and global annotations. First, a logarithmic transform-based guided spatial filter (LT-GSF) is applied to reduce speckle noise while keeping important edges clear. After preprocessing, a lightweight convolutional improved coordinate attention-based partial U-Net (LC-CAU-Net) is developed to accurately detect and segment regions of ovarian tumors. For classification, a knowledge distillation-based target-aware variational encoder (KD-TVE) is introduced to classify different tumor types with reduced complexity. Then, the Sine Chaos Parrot Optimization technique is used to tune hyperparameters for better performance and improved generalization. Experimental results show that the proposed LC-CAU-Net achieves the best Dice score of 96.77% and an IoU of 95.39%. The classification model reaches a highest accuracy of 98.17% and a best precision of 98.21%, outperforming existing methods. Overall, the proposed framework shows strong potential to support gynecologists in improving clinical diagnosis and treatment.
Indu et al. (Sat,) studied this question.