Deep Learning Based Junctional Zone Quantification using 3D Transvaginal Ultrasound in Assisted Reproduction

The Uterine Junctional Zone (JZ) is identified as an important anatomical region in the implantation process during assisted reproduction. The JZ changes throughout the hormone stimulation cycle and has predictive value for implantation success. Despite advances in imaging technique, the assessment of JZ remains an enigma. The state-of-the-art method to assess the JZ is largely manual, which is time consuming, depends on operator experience, and often introduces subjective bias in assessment. In this paper, we present methods for automated visualization and quantification of the JZ in three-dimensional transvaginal ultrasound imaging (3D-TVUS). JZ is best visualized in the midcoronal plane of the 3D-TVUS uterus acquisition. We propose an algorithm pipeline, which uses a deep learning model to generate a point cloud representing the surface of the endometrium. A regularized midcoronal surface passing through the point cloud is rendered to obtain the midcoronal plane. The automated solution is designed to accommodate multiple structural deformations and pathologies in the uterus. An expert assisted reproduction clinician on 136 3D-TVUS volumes evaluated the results, and reliable performance was observed in more than 89% cases where the automated solution is able to reproduce, and sometimes even outperform the manual workflow. Automation speeds up the clinical workflow approximately by a factor of ten and reduces operator bias.