Exploring Structural Information for Semantic Segmentation of Ultrasound Images

Biomedical semantic segmentation aims to automat-ically label each pixel of a medical image with a corresponding class, such as lung and heart. Such a technique significantly facilitates clinical diagnosis and thus has attracted increasing attentions both in industry and academia. Although promising, existing methods such as U-Net and MSRF, may suffer from the following two issues when they are applied to ultrasound images: 1) few ultrasound datasets are publicly available due to the regulation of patient's privacy, while U-NET/MSRF models are data-hungry and are trained on tens of thousands of images; 2) the low-quality ultrasonic images used in training procedure may lead to inaccurate segmentation in the evaluations. To fill this gap, this paper presents structured multi-scale residual fusion network (SMNet), a novel method for better semantic segmentation of ultrasonic images by exploring rich interactions between the organs. Equipped with such structured information, our SMNet can properly tackle the above two challenging issues. Experiments on a real-world dataset show the effectiveness of our approach, yielding significant improvement compared with the sate-of-the-art models.