Automatic C-Plane Detection in Pelvic Floor Transperineal Volumetric Ultrasound

Helena Williams, KU Leuven– University Hospital Leuven
Laura Cattani, KU Leuven– University Hospital Leuven
Mohammad yaqub, Mohamed Bin Zayed University of Artificial Intelligence
Carole Sudre, King's College London
Tom Vercauteren, King's College London
Jan Deprest, KU Leuven– University Hospital Leuven
Jan D’hooge, KU Leuven


Transperineal volumetric ultrasound (US) imaging has become routine practice for diagnosing pelvic floor disease (PFD). Hereto, clinical guidelines stipulate to make measurements in an anatomically defined 2D plane within a 3D volume, the so-called C-plane. This task is currently performed manually in clinical practice, which is labour-intensive and requires expert knowledge of pelvic floor anatomy, as no computer-aided C-plane method exists. To automate this process, we propose a novel, guideline-driven approach for automatic detection of the C-plane. The method uses a convolutional neural network (CNN) to identify extreme coordinates of the symphysis pubis and levator ani muscle (which define the C-plane) directly via landmark regression. The C-plane is identified in a postprocessing step. When evaluated on 100 US volumes, our best performing method (multi-task regression with UNet) achieved a mean error of 6.05 mm and 4.81 and took 20 s. Two experts blindly evaluated the quality of the automatically detected planes and manually defined the (gold standard) C-plane in terms of their clinical diagnostic quality. We show that the proposed method performs comparably to the manual definition. The automatic method reduces the average time to detect the C-plane by 100 s and reduces the need for high-level expertise in PFD US assessment.