Psoas MRI and muscle strength

1The orthopaedic team asked us whether one axial lumbar MRI slice at L4 to L5, together with a few patient facts, could tell us how strongly a muscle can pull.From Sudirman et al. 2019, CC BY 4.0

2We focused on the psoas because hip flexion depends on it, and its broad shape appears on each side of this axial image.From Sudirman et al. 2019, CC BY 4.0 · Model run on cuaim, job 406149

3We planned to test that idea against hip flexion force measured with a dynamometer, in newtons, so an image feature would face a physical reference.From Full Protocol, strength measurement · Model run on cuaim, job 406149

4To locate psoas consistently, we also separated the disc, spinal region and posterior muscle; those neighbouring regions gave the model anatomical boundaries to learn.From Preprocessing notebooks, annotation color keys · Model run on cuaim, job 406149

5Each training image carried painted regions, one colour per structure, while everything unpainted became background; we used those colours to define the segmentation targets.From Preprocessing notebooks, annotation color keys · Model run on cuaim, job 406149

6We first separated the paint by hue, then used image half to distinguish paired muscles, because a shared colour alone could not name each output channel.From Preprocessing notebooks, get_components · Model run on cuaim, job 406149

7We turned that preprocessing into one black and white target per class, so the network could learn eight mutually exclusive pixel categories from the painted training images.From 08_CNN_multi, input and target code · Model run on cuaim, job 406149

8We trained one U-Net on 256 by 256 inputs through four encoder levels and a 16 by 16 bottleneck, with eight channels, including the aorta and background.From 08_CNN_multi, model definition · Model run on cuaim, job 406149

9We concatenated each encoder level with its matching decoder level through four skip connections, preserving spatial detail while the eight softmax channels stayed on the original pixel grid.From 08_CNN_multi, model definition · Model run on cuaim, job 406149

10For our final outlines, we averaged two archived networks over flipped and rescaled copies of the slice; the aorta came from the 256 network's own probability channel.From 08_CNN_multi, model definition · 09_CNN_512_multi, model definition · Model run on cuaim, job 406149

11On this public MRI, our final outlines mark the psoas, disc, spinal region, back muscles and aorta, and every measurement that follows starts from these contours.From Sudirman et al. 2019, CC BY 4.0 · Model run on cuaim, job 406149

12Our later measurement code rotated each patch along a hand-drawn pink alignment guide, so that width and thickness followed the muscle's own orientation.From measurement/geometry.py · Model run on cuaim, job 406149

13We calculated minimum area rectangles from the model contours, then used the psoas region for area and its rotated box for width and thickness.From measurement/geometry.py · Model run on cuaim, job 406149

14We measured the corridor as the vertical separation between the rotated box tops; on this public slice they almost coincide, so the marked gap is small.From measurement/geometry.py · Model run on cuaim, job 406149

15We brought psoas area, rotated width and thickness, and basic patient facts toward a force model because anatomy alone may not capture muscle performance.From Full Protocol, objectives · prediction/run_predict.py · Model run on cuaim, job 406149

16Against that dynamometer reading in newtons, we ask whether psoas area, rotated width and thickness, and the patient facts predict the observed pull.From Full Protocol, strength measurement · prediction/run_predict.py · Model run on cuaim, job 406149

17Whether one MRI slice and a few facts can tell us how strong this muscle is remains the open question we built this work toward.From Full Protocol, objectives · Model run on cuaim, job 406149