TexNet: Texture Loss Based Network for Gastric Antrum Segmentation in Ultrasound

摘要

Gastric Antrum (GA) cross-sectional area measurement using ultrasound imaging is an important point-of-care (POC) application in intense care unit (ICU) and anesthesia. GA in ultrasound images often show substantial differences in both shape and texture among subjects, leading to a challenging task of automated segmentation. To the best of our knowledge, no work has been published for this task. Meanwhile, dice similarity coefficient (DSC) based loss function has been widely used by CNN-based segmentation methods. Simply calculating mask overlap, DSC is often biased towards shape and lack of generalization ability for cases with diversified and complicated texture patterns. In this paper, we present a robust segmentation method (TexNet) by introducing a new loss function based on multiscale information of local boundary texture. The new texture loss provides a complementary measure of texture-wise accuracy in contour area which can reduce overfitting issues caused by using DSC loss alone. Experiments have been performed on 8487 images from 121 patients. Results show that TexNet outperforms state of the art methods with higher accuracy and better consistency. Besides GA, the proposed method could potentially be an ideal solution to segment other organs with large variation in both shape and texture among subjects.