Deep learning-based thigh muscle segmentation for reproducible fat fraction quantification using fat–water decomposition MRI

Abstract

Background Time-efficient and accurate whole volume thigh muscle segmentation is a major challenge in moving from qualitative assessment of thigh muscle MRI to more quantitative methods. This study developed an automated whole thigh muscle segmentation method using deep learning for reproducible fat fraction quantification on fat–water decomposition MRI. Results This study was performed using a public reference database (Dataset 1, 25 scans) and a local clinical dataset (Dataset 2, 21 scans). A U-net was trained using 23 scans (16 from Dataset 1, seven from Dataset 2) to automatically segment four functional muscle groups: quadriceps femoris, sartorius, gracilis and hamstring. The segmentation accuracy was evaluated on an independent testing set (3 × 3 repeated scans in Dataset 1 and four scans in Dataset 2). The average Dice coefficients between manual and automated segmentation were > 0.85. The average percent difference (absolute) in volume was 7.57%, and the average difference (absolute) in mean fat fraction (meanFF) was 0.17%. The reproducibility in meanFF was calculated using intraclass correlation coefficients (ICCs) for the repeated scans, and automated segmentation produced overall higher ICCs than manual segmentation (0.921 vs. 0.902). A preliminary quantitative analysis was performed using two-sample t test to detect possible differences in meanFF between 14 normal and 14 abnormal (with fat infiltration) thighs in Dataset 2 using automated segmentation, and significantly higher meanFF was detected in abnormal thighs. Conclusions This automated thigh muscle segmentation exhibits excellent accuracy and higher reproducibility in fat fraction estimation compared to manual segmentation, which can be further used for quantifying fat infiltration in thigh muscles.

Automatic quadriceps and patellae segmentation of MRI with cascaded U2‐Net and SASSNet deep learning model

“…Moreover, imbalanced medical image data and high variability of target object shapes and locations often lead to unexpected segmentation results 57 . The anatomic symmetry and low distribution errors were ignored in previous models by a priori cropping 1,3,4,17–22,25,27 . The second refinement stage alleviates the memory bottleneck by focusing only on the cropped high‐resolution volume, provided by the low‐resolution first stage.…”

“…In the first stage, we localize the individual region of interest (VL, VM, VI, RF, or PA) as a coarse segmentation from the entire 3D MR image set. We first downsample the original images into the low‐resolution image space 27 . We utilize the multifeature images (anisotropic diffusion, coherence enhanced diffusion, regularized diffusion, gradient magnitude) throughout the two‐stage segmentation pipeline to stabilize the performance 5 .…”

Artificial intelligence for MRI diagnosis of joints: a scoping review of the current state-of-the-art of deep learning-based approaches

“…Similarly, deep learning-based segmentation has been described for thigh muscles and wrist bones [55,56].…”

Two decades of advances in muscle imaging in children: from pattern recognition of muscle diseases to quantification and machine learning approaches

“…Applications of DL in muscle MRI are rapidly expanding. Beyond segmentation [83] and automatic quantification [84] , some authors have developed models for dystrophynopathy recognition [85] and for assessment of the progressive course in collagen-VI related muscle disorders [86] .…”