U-Net Based Approach for Precise Ovarian Tumour Segmentation in Ultrasound Imaging
Ratnala Mani Deepika, P. K. A. Chitra · 2025
This paper proposes a deep learning-based U-Net model for the automated segmentation of ovarian tumours in ultrasound images, aiming to enhance diagnostic accuracy and support clinical decision-making. The model is trained and evaluated using the MMOTU: Multi-Modality Ovarian Tumour Ultrasound Image Dataset. To improve segmentation performance, Binary Cross-Entropy with Intersection over Union (BCE-IoU) loss is used during training. The proposed model achieves an Intersection over Union (IoU) score of 0.7592, a Dice coefficient of 0.8234, and a pixel-wise accuracy of 91.5 %, demonstrating competitive results when compared to existing methods such as ResNet-based segmentation, DeepLabV3+, and Attention U-Net. These outcomes highlight the effectiveness of the U-Net architecture in segmenting tumour regions from noisy and low-contrast ultrasound images. The model shows potential for integration into computer-aided diagnosis (CAD) systems to assist radiologists. Future enhancements will focus on incorporating attention mechanisms, leveraging multi-modal imaging data, and optimizing the model for real-time clinical use.