Table 1.
Variability of muscle volumes.
Fig 1.
Homogenisation of the fat tissue surrounding the muscles.
The process of masking the fat tissue surrounding the muscles from the raw MR images (left) and wrapping in a homogenous layer of fat for two images taken from different subjects (right). The subject along the top row (right) had a fat layer less that 5 mm thick and was wrapped with artificial fat, where the subject along the bottom row had a depth that was sufficient.
Fig 2.
Registration, multi-atlas, and image augmentation pipelines.
The image registration process, shown for one 2D slice of imaging data (location within imaging sequences highlighted with a black line). Segmentation pipeline: the target and reference subject were pre-processed, homogenizing the fat layer, and registered in ShIRT. The map found through registration was applied to the manual segmentation contours of the reference subject (shown in green), resulting in an automatic segmentation of the target subject (shown in blue). Data augmentation pipeline: The combined MR imaging sequences are registered in ShIRT. The map outputted from the registration was used to deform the reference subject’s 3D imaging data and reference manual segmentations, resulting in a fully segmented, augmented 3D image. The augmented images are shown with each muscle taking a different greyscale value (visualised in blue image channel). The multi-atlas pipeline used is also shown on the right-hand side. The outputs of the segmentation pipeline are combined forming a probability map for each muscle. The disputed and agreed voxels are separated, and the disputed voxels are labelled according to the best performing registration, found through calculating the mutual information of registered and fixed images. Collating these results provides a multi-atlas segmentation for each muscle.
Fig 3.
Sample registration outputs for qualitative interpretation.
Registration results of images taken from the shank. The registration quality is visualised within these plots with poor, moderate, and flawless registrations shown in a, b, and c, respectively. Yellow colour represents well registered regions.
Fig 4.
Qualitative interpretation of segmentation results.
Registration and segmentation results from the combination of subjects resulting in the median average DSC (subject 4 and 2 as the target and reference, respectively). The registration inputs (top row) and outputs (bottom row) for these combinations of subjects are shown on the group of images on the left. The segmentation results are shown in the right three image groups, where the reference and automatic segmentations for the target subject are shown in blue and red respectively. The muscles that are not highlighted within the images, were found not to be segmented with an appropriate level of repeatability.
Fig 5.
Numerical results for the intra-subject, multi-atlas, and single atlas analyses.
Relative volume error (%) (top), Dice similarity coefficient (centre), and Hausdorff distances (mm) (bottom) found for each muscle in each subject, for the left to right analysis (green), multi-atlas (orange), and inter-subject approaches (blue). In both the left to right and multi-atlas analysis, the numbers above the boxplots denote the subject segmented. The numbers above each of the boxplots in the inter-subject approaches denote the reference subject used within the registration. The dashed line in RVE plot shows the acceptable level of RVE resulting from inter-operator dependence, prescribed by Montefiori et al. 2020 [15]. The grey dashed lines in the DSC and HD plots represent the mean values from the intra-subject analysis for comparison. The box and whisker plots show the mean, interquartile ranges, and ranges across the 23 muscles considered.
Fig 6.
Inputs, outputs and resulting augmented subjects. Each row of images presents results within the hip (left), thigh (centre), and shank (right) for 3 subject combinations chosen at random (target x reference: 4 x 5 (top), 1 x 3 (middle), 7 x 9 (bottom)). Within each cell there are the inputted images into the registration (left), registered images with corresponding target image (centre) and resulting segmented, augmented images (right). The muscle labels are visible within the augmented images as the blue areas. Each muscle is assigned a distinct greyscale value and the labels are assigned alphabetically.
Fig 7.
Enhancement of muscle volume variability through image augmentation.
The anatomical variability of muscle volumes for each muscle, ordered from smallest to largest within the original and augmented databases shown in red and blue respectively. The height of the distributions was not normalised, and the violin plot contains 95% of the data, with 2.5% of data cut off from each side, removing outliers.