10th International Congress on Information and Communication Technology

Authors - Salimah Saeid, Tahani Almabruk, Muetaz Abdulsamad
Abstract - The diagnosis of ovarian tumors remains a challenging task due to the inherent variability and complexity of imaging data. This study evaluates the efficacy of transfer learning and fine-tuning techniques in convolutional neural net-works to enhance the classification accuracy of ovarian tumors in ultrasound images. The performance of YOLOv8 and VGG16 models were compared, including a modified VGG16 architecture optimized for this application. YOLOv8 models were evaluated both from scratch and with pre-trained weights, while VGG16 was employed for feature extraction and fine-tuning. The Modified VGG16 outperformed all other models, achieving the highest classification accuracy (%90) and the shortest training time (8.63 hours). Advanced data augmentation strategies and architectural optimizations effectively addressed issues such as class imbalance and overfitting. These results highlight the potential of customized CNN architectures and transfer learning to improve diagnostic accuracy and efficiency, advancing the development of reliable tools for ovarian tumor classification in clinical imaging.