Auto-context fully convolutional network for levator hiatus segmentation in ultrasoudn images

The accurate boundary segmentation of levator hiatus (LH) in pelvic ultrasound (PUS) images benefits the diagnosis of female pelvic floor dysfunction (FPFD). Since manual boundary delineation is tedious, time-consuming, and of large observer variability, automatic method is highly demanded to improve the efficacy and objectiveness of LH segmentation. In this paper, we propose an auto-context fully convolutional network (AC-FCN), in which a customized FCN is integrated with the auto-context model for iteratively refining the prediction map of LH. Benefited from the progressive refinement, the final prediction map possesses both the local spatial consistency and boundary details. Hence the LH boundary can be readily segmented by applying an auxiliary active shape model (ASM) on this prediction map. The efficacy of the proposed method is validated on extensive challenging LH PUS images. Experimental results demonstrate that the proposed method provides more accurate segmentation than state-of-the-art methods do.