Domain‐specific data augmentation for segmenting MR images of fatty infiltrated human thighs with neural networks

Abstract

Background Fat‐fraction has been established as a relevant marker for the assessment and diagnosis of neuromuscular diseases. For computing this metric, segmentation of muscle tissue in MR images is a first crucial step. Purpose To tackle the high degree of variability in combination with the high annotation effort for training supervised segmentation models (such as fully convolutional neural networks). Study Type Prospective. Subjects In all, 41 patients consisting of 20 patients showing fatty infiltration and 21 healthy subjects. Field Strength/Sequence: The T1‐weighted MR‐pulse sequences were acquired on a 1.5T scanner. Assessment To increase performance with limited training data, we propose a domain‐specific technique for simulating fatty infiltrations (i.e., texture augmentation) in nonaffected subjects' MR images in combination with shape augmentation. For simulating the fatty infiltrations, we make use of an architecture comprising several competing networks (generative adversarial networks) that facilitate a realistic artificial conversion between healthy and infiltrated MR images. Finally, we assess the segmentation accuracy (Dice similarity coefficient). Statistical Tests A Wilcoxon signed rank test was performed to assess whether differences in segmentation accuracy are significant. Results The mean Dice similarity coefficients significantly increase from 0.84–0.88 (P < 0.01) using data augmentation if training is performed with mixed data and from 0.59–0.87 (P < 0.001) if training is conducted with healthy subjects only. Data Conclusion Domain‐specific data adaptation is highly suitable for facilitating neural network‐based segmentation of thighs with feasible manual effort for creating training data. The results even suggest an approach completely bypassing manual annotations. Level of Evidence: 4 Technical Efficacy: Stage 3

Overview of MR Image Segmentation Strategies in Neuromuscular Disorders

“…However, such a classification did not allow to distinguish perimuscular and intramuscular adipose tissue. Using a U-Net architecture, Gadermayr et al intended to segment healthy and fat-infiltrated muscle all-together on T 1weighted images, allowing the distinction of intramuscular from perimuscular adipose tissue (68). Given the complexity of this task, corresponding DSC values were smaller (around 0.88 ± 0.05), illustrating a poorer segmentation quality.…”

“…Recently, Yi et al (78) made a review regarding the application of such methods in medical imaging. For lower limb muscle segmentation, solutions based on GAN were assessed with the aim of generating pathological images (68). Many issues related to the realistic nature of the generated images and their variability have still to be addressed.…”

Supervised segmentation framework for evaluation of diffusion tensor imaging indices in skeletal muscle

“…12 Therefore, there is great need for (semi-) automatic segmentation tools in the context of quantitative imaging of skeletal muscle. Recently, some (semi-) automatic approaches have been proposed for skeletal muscle [12][13][14][15][16][17][18][19][20] and showed good correspondence, reflected by high Dice similarity coefficient (DSC) values, with manual segmentation. However, these approaches only focused on either a partial volume of the muscles or entire muscle groups rather than the full volume of individual muscles.…”

DuCaGAN: Unified Dual Capsule Generative Adversarial Network for Unsupervised Image-to-Image Translation

“…For example, due to the time-consumption and high cost for creating a large amount of paired data for autonomous driving is timeconsuming and costly, the image-to-image translation method is used to enrich the dataset of the autonomous driving scenes by synthesizing various street scene images to improve the learning ability [2], [3]. Especially, image augmentation applied to surface defect data from the real industrial field can also be formalized as the image translation problem [4], [5].…”

“…To enrich the defect sample images, the traditional image processing methods, such as copy, rotating, and cropping, are employed, while they can not display the defect features correctly. Therefore, insufficient sample sizes and sample class imbalances have become an urgent problem to be solved for the defect data in the real industrial process [4], [5]. Although some previous image translation methods had not focused on solving this problem, we can transfer it into an unsupervised image translation problem by respectively modeling samples as the normal domain and the defect domain, respectively.…”