Deep learning methods for automatic segmentation of lower leg muscles and bones from MRI scans of children with and without cerebral palsy

Abstract

Cerebral palsy is a neurological condition that is known to affect muscle growth.Detailed investigations of muscle growth require segmentation of muscles from MRI scans, which is typically done manually. In this study, we evaluated the performance of 2D, 3D, and hybrid deep learning models for automatic segmentation of 11 lower leg muscles and two bones from MRI scans of children with and without cerebral palsy. All six models were trained and evaluated on manually segmented T 1 -weighted MRI scans of the lower legs of 20 children, six of whom had cerebral palsy. The segmentation results were assessed using the median Dice similarity coefficient (DSC), average symmetric surface distance (ASSD), and volume error (VError) of all 13 labels of every scan. The best performance was achieved by H-DenseUNet, a hybrid model (DSC 0.90, ASSD 0.5 mm, and VError 2.6 cm 3 ). The performance was equivalent to the inter-rater performance of manual segmentation (DSC 0.89, ASSD 0.6 mm, and VError 3.3 cm 3 ). Models trained with the Dice loss function outperformed models trained with the cross-entropy loss function. Near-optimal performance could be attained using only 11 scans for training. Segmentation performance was similar for scans of typically developing children (DSC 0.90, ASSD 0.5 mm, and VError 2.8 cm 3 ) and children with cerebral palsy (DSC 0.85, ASSD 0.6 mm, and VError 2.4 cm 3 ). These findings demonstrate the feasibility of fully automatic segmentation of individual muscles and bones from MRI scans of children with and without cerebral palsy.

Deformable image registration for automatic muscle segmentation and the generation of augmented imaging datasets

“…The geometry of the 23 muscles was captured moderately well only in the optimal combination of subjects (those with greatest lower quartile), with mean DSC of 0.74 and IQR range of 0.71 < DSC < 0.77. However, this is significantly smaller than the inter-operator dependence of the manual process, which, within the literature [2,15,21,24] is consistently found to be DSCs of around 0.90 for the muscles considered in this study. While the pair of subjects leading to the best results in terms of DSC were the most similar in terms of height and BMI, these anthropometric characteristics were very different in the pair having the second-best DSC (mean = 0.74, IQR of 0.69 < DSC < 0.79).…”

“…Other methods have been shown to be effective in the segmentation of muscles. Probabilistic machine learning methods such as deep learning have been used to automatically segment the 3D geometry of individual muscles from MR images taken from several different cohorts [20,21,24]. These methods employ Convolutional Neural Networks (CNNs) which learn patterns that identify important features from training data in order to apply these learned patterns to segment new, unseen data [20,21,24].…”

“…Probabilistic machine learning methods such as deep learning have been used to automatically segment the 3D geometry of individual muscles from MR images taken from several different cohorts [20,21,24]. These methods employ Convolutional Neural Networks (CNNs) which learn patterns that identify important features from training data in order to apply these learned patterns to segment new, unseen data [20,21,24]. Notably, the methods recently proposed by Ni et al [24], where all lower limb muscles within a cohort (n = 64) of young healthy athletes were segmented with DSC comparable to that of the inter-operator dependence (DSC ≈ 0.9), and those proposed by Zhu et al [21] where all muscles within the shank were segmented from a cohort (n = 20) of children with cerebral palsy (DSC ≈ 0.88).…”

“…Not only is this interaction time excessive, but operators must undergo training to achieve repeatable segmentation results from an intra-operator standpoint ( ± 10% volume is typically acceptable) [15]. Regardless of training, as suggested, there are significant inter-operator dependency issues noted within the literature, which have been shown to misinterpret muscle volume by up to 50% (for example, the peroneus brevis and longus [15]), depending on the muscle of interest and study cohort [15,21,24]. These limitations of manual segmentation prevent the utilisation of muscle segmentation as a technique to inform large-scale quantitative investigations into muscle characteristics.…”

“…There are two main limitations of manual segmentation: the required operator input time and operator dependency issues of the outputs [5,6,15,18]. It is generally accepted that there are 35 individual lower limb muscles, not including those in the feet (some of these muscles can be separated into different branches or sub-muscles [15,22,23]), which must be segmented from (in the order of) hundreds of images, incurring a high processing time [15,20,21]. Recent advancements in computer vision (interpolation between segmented slices) and hardware (trackpads) have brought operator interaction time down to approximately 10 hours to segment all muscles within one lower limb [15,20,21].…”