Deformable image registration based on single or multi-atlas methods for automatic muscle segmentation and the generation of augmented imaging datasets

Abstract

Muscle segmentation is a process relied upon to gather medical image-based muscle characterisation, useful in directly assessing muscle volume and geometry, that can be used as inputs to musculoskeletal modelling pipelines. Manual or semi-automatic techniques are typically employed to segment the muscles and quantify their properties, but they require significant manual labour and incur operator repeatability issues. In this study an automatic process is presented, aiming to segment all lower limb muscles from Magnetic Resonance (MR) imaging data simultaneously using three-dimensional (3D) deformable image registration (single inputs or multi-atlas). Twenty-three of the major lower limb skeletal muscles were segmented from five subjects, with an average Dice similarity coefficient of 0.72, and average absolute relative volume error (RVE) of 12.7% (average relative volume error of -2.2%) considering the optimal subject combinations. The multi-atlas approach showed slightly better accuracy (average DSC: 0.73; average RVE: 1.67%). Segmented MR imaging datasets of the lower limb are not widely available in the literature, limiting the potential of new, probabilistic methods such as deep learning to be used in the context of muscle segmentation. In this work, Non-linear deformable image registration is used to generate 69 manually checked, segmented, 3D, artificial datasets, allowing access for future studies to use these new methods, with a large amount of reliable reference data.

Automatic segmentation of lower limb muscles from MR images of post-menopausal women based on deep learning and data augmentation

“…In this case, the spatial channel would prevent the inclusion of this muscle in the prediction for that certain image, which could lead to an incorrect segmentation. As is clear from Fig 6 , when retrained with additional augmented images, which were shown to increase the variability of muscle structure when comparing the augmented and original datasets [ 25 ], this effect was reduced for the SC-UNet. Nevertheless, it remains to be tested if a further augmentation of the dataset would lead to even higher improvement of the evaluation metrics for the SC-UNet.…”

“…The CNNs were trained using both the original MRI and a database of augmented images. These augmented images were generated in a previous study [ 25 ], using deformable image registration. Each of the 11 subjects were paired with one another and their associated MRI were input to the registration pipeline, using both subjects as both the fixed and the moving image.…”

“…Each of the 11 subjects were registered to one another, generating 110 anatomically distinct augmented imaging datasets. As described in [ 25 ], these subjects were vetted by experts in muscle segmentation according to a set of exclusion criteria requiring that the augmented subjects were anatomically feasible, after which the number of suitable 3D imaging datasets (or virtual subjects) of the lower limb was reduced to 69. In a previous study [ 34 ], the anatomical variability of the augmented database was compared to that of the original database, highlighting that the shape and size of the muscles within the augmented database were significantly different to those within the original database.…”

“…The CNNs were retrained including the augmented images (generated through deformable image registration) [ 25 ] in the training dataset. The same subject as in test 1 was segmented allowing the effect of including the augmented images to be analysed.…”

“…Nevertheless, muscle segmentation using CNN approaches applied to MRI is challenging because there are few openly available datasets that would be large enough to facilitate learning-based methods, and that the available datasets are inhomogeneous in their acquisition parameters [ 23 ]. Other segmentation techniques using deformable registration and statistical shape modelling have been explored but showed only limited accuracy (DSC in the range of 0.90–0.99 for groups of muscles, and 0.40–0.87 for single muscles) can be achieved [ 24 , 25 ]. For example, in a previous study, the use of a single and multi-atlas image registration algorithm (the Sheffield Image Registration Toolkit, or ShIRT [ 26 ]) was explored for automatic muscle segmentation of a subset (N = 5) of the study cohort used in this paper.…”