Automatic muscle segmentation on healthy abdominal MRI using nnUNet

Victoria Joppin, Niamh Belton, Marc Adrien Hostin, Marc-Emmanuel Bellemare, Aonghus Lawlor, Kathleen M Curran, Thierry Bège, Catherine Masson, David Bendahan

Keywords: Medical imaging segmentation, Deep learning, Abdominal wall muscles

Abstract:

Understanding the dynamics of the abdominal wall is essential in both physiology and
surgery. To study the mechanical functionality of the abdominal wall, segmentation of the
abdominal muscles could be useful but is a manual, tedious and time-consuming process.
In this study, we assessed the capability of Deep Learning to automatically segment the
abdominal muscles from the axial plane of a unique dynamic abdominal MRI (2D+t)
dataset. The 2D slices were acquired while the subject performed various exercises. The
State-of-the-Art segmentation model, nnUNet was trained on 5, 492 images from fifteen
healthy subjects and tested it on 1, 801 images from five different subjects. We assessed
the segmentation accuracy using DICE similarity coefficient, Hausdorff distance, as well as
motion of the abdominal muscles. The ground truth and nnUNet segmentation showed high
concordance, with a DICE over 0.87 for all exercises and muscles, and minimal differences in
abdominal muscles motion. nnUNet effectively automates abdominal muscle segmentation,
offering efficiency and new clinical applications in abdominal physiology.

Submission Number: 175