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.
- Transperineal volumetric ultrasound,
- pelvic floor disease,
- convolutional neural network,
- plane detection