Fully-automated segmentation of muscle and inter-/intra-muscular fat from magnetic resonance images of calves and thighs: an open-source workflow in Python

Figures

• 

  Fig. 1. Sources of precision error in the previously-validated semi-automatic application of the ITSA [14]. (A) An example of a raw axial FSE MR slice in the mid-calf. (B) Zoomed in view of A showing distribution of brighter and fainter (partial-volumed) voxels of fat. (C) Successful IMF segmentation applied by ITSA (in yellow), showing remaining voxels of fat that could be quantified. The white arrow points to remaining fat that was not captured due to its greyscale values being below the final threshold identified by ITSA. (D) Histogram showing distribution of fat pixel signal intensities

• 

  Fig. 2. Automation of muscle-fat ROI on which ITSA was applied. (A) The Otsu algorithm was used to estimate a threshold that can be used to generate a binary mask of the muscle border. The OpenCV contours function is used to determine the coordinates of the estimated muscle border, represented by the red dashed line. (B) Snake algorithm is used to adjust and improve accuracy of the muscle border coordinates, represented by the blue dashed line. Mistakenly included a piece of subcutaneous fat in contact with the tibia, and marked for removal is in orange. (C) Bone, represented by grey, is segmented by expanding a seed point identified by maximum summated marrow signal across slices, merged with a void signal mask containing cortices and removed from the ROI, represented by white. (D) Final muscle mask represented by translucent red is overlaid on top of the raw MRI slice. (E) final cropped muscle ROI

• 

  Fig. 3. Exported tags for manual correction within Sliceomatic software

• 

  Fig. 4. Representative image of partial volumed voxels of fat meriting a second round of ITSA. (A) Original MR image. (B) Yellow showing IMF segmentation result from applying a first round of ITSA, and remaining light grey areas suggesting outstanding partial volumed fat voxels. (C) Red showing IMF segmentation result from applying a second round of ITSA. (D) Final segmentation in full view

• 

  Fig. 5. Reclassification of islands of IMF voxels. (A) Single slice view of IMF segmentation from round 1 and 2 of ITSA. (B) 3-dimensional plot of fat voxels within a sample cropped 30 × 30 × 10 matrix. Orange depicts definite fat, > = 16 mm2 in the X–Y plane. Grey depicts unconnected noise, < 8 voxels. Green depicts islands of fat < 16 mm2 but connected to fat across adjacent slice(s). (C-E) orthogonal views of 3D model in the X–Y, X–Z, and Z-Y planes, respectively

• 

  Fig. 6. Examples of (A) IMF merging with subcutaneous fat, (B) motion streaks over-segmenting fat, and (C) noise within lean muscle segmented as fat

• 

  Fig. 7. Automation of ITSA IMF segmentation for a prototypical leg (A-C) versus an individual with muscle dystrophy (D-F). (A & D) Original image. (B & E) Isolation of muscle ROI after filters, snake and convex hull algorithms. (C & F) Resultant IMF auto-segmentation without corrections displaying yellow as fat identified in round 1 of ITSA and red as partial volumed fat from a second round of ITSA

• 

  Fig. 8. Bland–Altman analysis for differences in (A) IMF volume, (B) IMF%, and (C) Muscle volume metrics between the fully-automated versus semi-automated methods. Red dotted lines indicate 95% limits of agreement. Solid lines indicate mean difference between methods

• 

  Fig. 9. Scatter plots illustrating validity of IMF and muscle volumetric measures. Dotted lines represent univariable fitted regression slopes. Unity (1.0) indicates an ideal fit. Values < 1.0 indicate fully-automated method yielding smaller values than semi-automated method. Values > 1.0 indicate fully-automated methods yielding larger values than semi-automated method