Marker-Aware Ovarian Tumor Segmentation from Ultrasound Images

Ovarian cancer remains one of the leading causes of cancer-related deaths among women, with early detection being pivotal for successful treatment. Accurate segmentation of ovarian tumor regions in ultrasound images is essential to assist clinicians in the effective diagnosis and treatment of ovarian cancer. However, the complex nature of ultrasound images, with their inherent noise and the symbols marked by sonographers, poses significant challenges for ovarian tumor segmentation. In this study, before feeding the images into segmentation model, we employ an in-painting method to remove the symbols and markers from ultrasound image. Then a segmentation model based on UNet3+ with ResNet50 as its encoder is introduced. The results of our model on OTU2D dataset demonstrate an improvement over the existing models with the Dice, IoU, Recall and Precision scores of 86.44%, 77.05%, 86.31% and 89.18%. Moreover, when testing on a subset of clean images, the preprocessing technique based on in-painting helps to increase the Recall and Precision metrics from 89.32% and 86.23% to 97.78% and 96.08%. This study offers a promising solution to enhance image segmentation capabilities and aid clinicians in making better-informed clinical decisions.