Efficient data labeling strategies for automated muscle segmentation in lower leg MRIs of Charcot-Marie-Tooth disease patients

Seung-Ah Lee; Hyun Su Kim; Ehwa Yang; Young Cheol Yoon; Ji Hyun Lee; Byung-Ok Choi; Jae-Hun Kim

doi:10.1371/journal.pone.0310203

Abstract

We aimed to develop efficient data labeling strategies for ground truth segmentation in lower-leg magnetic resonance imaging (MRI) of patients with Charcot-Marie-Tooth disease (CMT) and to develop an automated muscle segmentation model using different labeling approaches. The impact of using unlabeled data on model performance was further examined. Using axial T1-weighted MRIs of 120 patients with CMT (60 each with mild and severe intramuscular fat infiltration), we compared the performance of segmentation models obtained using several different labeling strategies. The effect of leveraging unlabeled data on segmentation performance was evaluated by comparing the performances of few-supervised, semi-supervised (mean teacher model), and fully-supervised learning models. We employed a 2D U-Net architecture and assessed its performance by comparing the average Dice coefficients (ADC) using paired t-tests with Bonferroni correction. Among few-supervised models utilizing 10% labeled data, labeling three slices (the uppermost, central, and lowermost slices) per subject exhibited a significantly higher ADC (90.84±3.46%) compared with other strategies using a single image slice per subject (uppermost, 87.79±4.41%; central, 89.42±4.07%; lowermost, 89.29±4.71%, p < 0.0001) or all slices per subject (85.97±9.82%, p < 0.0001). Moreover, semi-supervised learning significantly enhanced the segmentation performance. The semi-supervised model using the three-slices strategy showed the highest segmentation performance (91.03±3.67%) among 10% labeled set models. Fully-supervised model showed an ADC of 91.39±3.76. A three-slice-based labeling strategy for ground truth segmentation is the most efficient method for developing automated muscle segmentation models of CMT lower leg MRI. Additionally, semi-supervised learning with unlabeled data significantly enhances segmentation performance.

Citation: Lee S-A, Kim HS, Yang E, Yoon YC, Lee JH, Choi B-O, et al. (2024) Efficient data labeling strategies for automated muscle segmentation in lower leg MRIs of Charcot-Marie-Tooth disease patients. PLoS ONE 19(9): e0310203. https://doi.org/10.1371/journal.pone.0310203

Editor: Claudia Brogna, Fondazione Policlinico Universitario Gemelli IRCCS, ITALY

Received: June 13, 2024; Accepted: August 24, 2024; Published: September 6, 2024

Copyright: © 2024 Lee et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Data Availability: The data generated during the current study are provided in S2 File.

Funding: This research was supported by National Research Foundation of Korea (NRF) grants funded by the Korean government (NRF-2022R1F1A1064070 and NRF-2022R1A2C1004516). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Competing interests: The authors have declared that no competing interests exist.

Abbreviations: CMT, Charcot-Marie-Tooth disease; MRI, magnetic resonance imaging; ADC, average dice coefficient; EMA, exponential moving average

Introduction

Charcot–Marie–Tooth disease (CMT) is the most common hereditary neuromuscular disorder globally and is associated with numerous gene mutations linked to axonal degeneration [1]. CMT is characterized by distal-predominant limb muscle wasting and sensory loss [1], making careful assessment and monitoring of affected muscles crucial for proper patient management. Magnetic resonance imaging (MRI) plays a significant role in the evaluation of patients with CMT, allowing an objective assessment of limb muscle volume and quality [2–5]. Muscle denervation leads to the degeneration of fibers and their replacement by intramuscular fat cells, identifiable as volume loss and increased intramuscular fat extent on MRI [6, 7]. MRI biomarkers, including muscle volume, fat fraction, and pattern of muscle denervation with specific regard to muscle compartments in the lower legs, are clinically important for prognosis estimation, clinical assessment, and identification of certain genetic types in patients [8–12].

Extracting MRI biomarkers necessitates muscle segmentation with regions of interest (ROI) lying within the muscle compartments, excluding other structures such as the subcutaneous fat tissue or bones. However, manual segmentation is both time-consuming and labor-intensive, as well as subject to intra- and interobserver variances [13]. Several deep-learning-based models have recently emerged as promising tools for automated muscle segmentation, which could provide effective solutions for MRI analyses that currently rely on manual segmentation [7, 14–18]. However, developing such automated segmentation requires access to training data comprising manually segmented images [7, 14]. Muscle segmentation requires meticulous handling to define the boundary between the muscles and subcutaneous fat tissue, particularly in CMT patients with intramuscular fat infiltration and volume loss on MRI [3]. Thus, manual ground truth segmentation of MRIs scans of patients with CMT can be substantially labor-intensive, and data on the most efficient method for this process are limited.

Although there is currently no cure for CMT, numerous clinical and preclinical studies involving therapeutic agents have shown promising results [19–21]. Consequently, it is imperative to develop and establish an objective and reproducible method for evaluating patients with CMT, including assessing their response to treatment. To our knowledge, most previous studies have focused on developing automated segmentation models for thigh muscles on MRI [7, 14, 22, 23]. This focus might be attributed to the less complex anatomy and larger muscle volume of the thighs, making segmentation less challenging compared to that of the lower legs. However, lower leg muscles are more severely affected in CMT due to the characteristic distal-predominant limb muscle wasting. Therefore, it would be more beneficial to develop a model that can accurately segment muscles in the lower legs. In this context, we assume that the development of a robust and reliable automated muscle segmentation model for the evaluation of lower leg muscles on MRI represents a challenging yet highly desirable long-term objective.

Herein, we investigated the most efficient data labeling strategy for ground truth segmentation of the lower leg muscles in the MRIs of patients with CMT to develop an automated muscle segmentation model. We compared the performance of the segmentation models obtained using several different labeling strategies for the training data. Additionally, we investigated the effect of utilizing unlabeled data to enhance muscle segmentation performance by comparing the fully-supervised, few-supervised, and semi-supervised learning models.

Materials and methods

Patient selection

This study was approved by the Institutional Review Board of our institution (Samsung medical center, IRB File No. 2022-11-073-001), and the requirement for informed consent was waived owing to its retrospective nature. Further, this study was conducted in accordance with the principles of the Declaration of Helsinki. Initially, we identified 273 patients diagnosed with CMT confirmed to have a PMP22 duplication through genetic analysis at our neurology department and who underwent lower extremity MRI between September 2013 and December 2022. A radiologist with 6 years of experience in musculoskeletal radiology (BLINDED) reviewed and categorized patients as either ’mild’ or ’severe’ based on the degree of lower leg muscle fat infiltration using the five-point semi-quantitative scale described by Goutallier et al. [15] Cases with Goutallier grade 3 or 4 fat infiltration on T1-weighted images, indicating 50% or higher fatty degeneration, were categorized as ’severe.’ The MRI scans of 120 patients (60 with mild and 60 with severe fat infiltration) were used as a training set. Data were initially accessed for research purpose on May 2^nd 2023. Authors did not have access to information that could identify individual participants during or after data collection.

MRI protocol

Images were obtained using a 3.0-T MRI system (Ingenia; Philips Healthcare, Best, Netherlands) with a 16-channel anterior coil and a posterior built-in coil. MRI protocols including the following sequences were obtained for the pelvic girdle, bilateral thighs, and lower legs to evaluate the lower extremity muscles: axial and coronal T1-weighted turbo spin echo sequences and axial T2 -weighted Dixon sequence. To acquire the axial T1-weighted turbo spin-echo sequence of the lower legs, the following parameters were used: repetition time, 612 ms; echo time, 11 ms; flip angle, 90°; number of signals averaged, 1; field of view, 284 × 379 mm; section thickness, 3 mm; gap, 11 mm; matrix size, 448 × 269; echo train length, 3; and number of slices, 32.

Ground truth segmentation

A radiologist (BLINDED) first selected the axial T1-weighted image of the proximal lower leg at a level just inferior to the inferior margin of the popliteus muscle, where the muscle was no longer visualized. The axial image of the distal lower leg was selected at the level of the uppermost part of the gastrocnemius tendon, where the gastrocnemius muscle was no longer visualized. The axial images selected for the proximal and distal lower leg, based on anatomical landmarks, have been used in numerous studies employing MRI to evaluate lower leg muscles in patients with CMT [9, 24–26]. These two axial images, along with others at intermediate levels, were used to analyze each patient. A single rater (BLINDED) manually delineated the boundaries of the bilateral lower leg muscle compartments on each slice using ITK-SNAP software (http://www.itksnap.org, accessed December 30, 2020) under a radiologist’s supervision. The muscle compartments were categorized as follows: anterior (tibialis anterior and extensor digitorum longus), lateral (peroneus longus), deep posterior (tibialis posterior), and superficial (gastrocnemius and soleus). Boundaries were drawn to avoid the extramuscular fat tissue, bony cortex, and neurovascular bundles.

Comparisons of models using different training data labeling strategies

We devised several strategies for labeling the MRI data for training models, considering inter-subject and inter-slice variations of the data. Intersubject variation is related to anatomical variation, as well as the degree of intramuscular fat infiltration in each subject, which is linked to the severity of denervation in patients with CMT. Interslice variation is related to variations in the substantially differing lower leg muscle compartment anatomy, depending on the proximal-to-distal level [27]. Specifically, we categorized the strategies as “single slice,” “three-slices,” and “all-slices strategies” (Fig 1). For the “single slice strategy,” a single image slice per subject was used to train the model to account for intersubject variation by maximizing the number of subjects that were included for training. The single image provided was chosen either as the uppermost, the central, or the lowermost axial T1-weighted images among segmented lower leg images, of which the strategy was named as “uppermost,” “central,” and “lowermost strategy,” respectively. The “three-slices strategy” utilized the uppermost, central, and lowermost image slices of each subject. The number of subjects used in this strategy was nearly one-third of that of the “single slice strategy,” given that the same amount of data were used for training. The “all-slices strategy” utilized all the axial images of the lower legs per subject. Thus, the number of subjects used in this strategy was smaller than that of the “three-slices strategy.” This strategy was devised to account for inter-slice variation by maximizing the number of proximal-to-distal levels of the axial images within the subjects.

Download:

Fig 1. Data labeling strategies for ground truth segmentation of the lower leg MRI data.

For “single slice strategy”, single image slice per subject was used to train the model to maximize the number of subjects included for training. The single image provided was chosen as either the uppermost, the central, or the lowermost axial T1-weighted images among segmented lower leg images, of which the strategy was named as “uppermost,” “central,” and “lowermost strategy,” respectively. The “three-slices strategy” utilized the uppermost, central, and lowermost image slices of each subject. The number of subjects used in this strategy was nearly one-third of that of the “single slice strategy,” given the same amount of data were used for training. The “all-slices strategy” utilized all the axial images of the lower legs per subject. Thus, the number of subjects used in this strategy was even smaller than that of the “three-slices strategy”.

https://doi.org/10.1371/journal.pone.0310203.g001

Comparison of models using different learning methods and different amount of training data

We compared the muscle segmentation performance of the fully-supervised learning model and the few- and semi-supervised learning models of the three different data labeling strategies (Fig 2). Fully-supervised learning employs all the labeled data for training, whereas few-supervised learning employs specific proportions (2%, 5%, and 10%) of the labeled data. The semisupervised learning model utilizes a combination of labeled and unlabeled data during training; this method has been reported to demonstrate results comparable to those of fully supervised learning [28, 29]. After model training, we compared the muscle segmentation performances of the fully supervised learning model and the few-supervised and semi-supervised learning models of three different data labeling strategies.

Download:

Fig 2. Outline of the research methodology.

We compared muscle segmentation performances of fully-supervised learning models and few-supervised and semi-supervised learning models of three different data labeling strategies. The fully-supervised learning employed all labeled data for training, while few-supervised learning employed specific proportions of the labeled data. The semi-supervised learning model utilized a combination of labeled and unlabeled data during training.

https://doi.org/10.1371/journal.pone.0310203.g002

Image preprocessing

MRIs were converted into the TensorFlow TFRecords format for dataset loading. The 3D images were separated into 2D images and were subsequently cropped to 448 × 672 pixels along the y-axis and resized to half of the cropped size to accelerate computation and the effect of data augmentation. The intensities of the MRIs scans were normalized using mean and standard deviation of each slice, standardizing pixel intensities across images to have a mean of 0 and a standard deviation of 1.

2D U-net: 2D segmentation model

We used a 2D U-Net to segment the four muscle compartments on MRIs, as U-Net has been shown to provide promising medical image segmentation performance [30, 31]. U-Net is a fully convolutional network that consists of an encoder, a bottleneck, and a decoder (Fig 3). The encoder downsamples the input image to extract high-level features, while the decoder upsamples these features to match the original image resolution. Skip connections between the encoder and decoder enable the model to access low-level semantic information from the encoder, which again facilitates the generation of desired features by the decoder (S1 File).

Download:

Fig 3. Mean teacher framework for semi-supervised model and the architecture of the U-net for lower leg muscle segmentation.

https://doi.org/10.1371/journal.pone.0310203.g003

Semi-supervised deep learning model

We used the mean teacher classification model for semi-supervised learning [32, 33]. The mean-teacher model uses two neural networks, a student network, and a teacher network, to train on both labeled and unlabeled data. During training in the supervised phase, the student network was updated with both labeled and unlabeled data using gradient descent, whereas the teacher network’s weights remained fixed. In the unsupervised phase, both networks predict unlabeled data, and the predictions of the student network were compared with those of the teacher network. The goal was to minimize the difference between the two predictions, which helped the student network learn from the unlabeled data. At each step, the student network’s weights are updated with the teacher network’s weights, called exponential moving average (EMA) weight sharing [32, 34]. The EMA stabilizes training, maintains consistency, and aids the student model in better generalization.

Implementation details

The dataset was divided into training and testing sets (80%:20%), assuming an equal number of labeled slices for each labeling strategy. We randomly sampled 2, 5%, and 10% of the labeled data as labeled sets and used the remaining labeled data as an unlabeled set (Table 1). The experiments were implemented using TensorFlow 2.6.0 and trained for 1000 epochs with a batch size 32 on two GTX 1080 Ti GPUs. Data augmentation using the albumentation library involved horizontal flipping, rotation within (-30, 30) degrees, scaling with a factor in the range (-0.5~0.3) for both height and width, and shifting with a factor in the range (-0.2, 0.2) [13].

Download:

Table 1. Number of images used for the training and testing of the models with different labeling strategies.

https://doi.org/10.1371/journal.pone.0310203.t001

The segmentation performance of each model was measured using the Dice coefficient [13], which is the spatial overlap between the predicted mask and ground truth: (1)

In Eq 1, TP represents true positive predictions, FP represents false positive predictions, and FN represents false negative predictions. During training, we utilized the Dice loss function (1-Dice) and Adam optimizer with a learning rate of 1e-3.

The mean squared error was employed as the consistency loss to minimize the difference between the predictions of the student and teacher models in unsupervised learning.

(2)

(3)

In semi-supervised learning, the hyper-parameters comprised the unsupervised loss weight factor λ, and the EMA decay rate α. The unsupervised loss weight factor γ followed a sigmoid curve from 0 to 1. The model updated teacher model weights θ′ at training step t using Eq 3, with the updating rate governed by the EMA decay parameter α set to 0.999.

Statistical analyses

We compared the performances of the segmentation models obtained using different labeling strategies by comparing the average Dice coefficients (ADC) of the four muscle compartments. First, pairwise comparisons were conducted among the few supervised learning models trained with a 10% labeled set with different labeling strategies (single slice [uppermost, central, or lowermost], three-slices, and all-slices). Pairwise comparisons between few-supervised and semi-supervised learning models of different strategies were performed to verify the benefits of the semi-supervised methods. Further comparisons were performed based on the proportion of labeled data used for training (2%, 5%, and 10%). Paired t-tests were employed, and the Bonferroni correction was applied for multiple comparisons. Statistical analyses were performed using SAS version 9.4 (SAS Institute, Cary, NC, USA), considering a P-value < 0.05 as statistically significant.

Results

Comparisons of models using different training data labeling strategies

ADCs of the few-supervised learning models trained with 10% labeled set with single slice strategies were uppermost, 87.79±4.41%; central, 89.42±4.07%; and lowermost, 89.29±4.71% (mean ± standard deviation). Pairwise comparisons among the models using different single-slice strategies revealed that the ADCs of the central and lowermost slice strategies were significantly higher than those of the uppermost slice strategy (p < 0.0001; Figs 4 and 5). The ADCs of the three-slices strategy and all-slices strategy were 90.84±3.46% and 85.97±9.82%, respectively. Pairwise comparisons between the two models showed that the model trained using the three-slices strategy had a significantly higher ADC than that trained using the all-slices strategy (p < 0.0001). The model using the three-slices strategy also showed a significantly higher ADC than that using the central-slice strategy (p < 0.0001).

Download:

Fig 4. Boxplots for the comparisons of average dice coefficients among segmentation models using different training data labeling strategies.

Each model was trained using 10% of the labeled data except for the fully-supervised model. * indicates statistical significance.

https://doi.org/10.1371/journal.pone.0310203.g004

Download:

Fig 5.

Examples of MRI, ground truth, and segmentation results of few-supervised learning models with 10% labeled data using different labeling strategies in mild (upper row) and severe (lower row) intramuscular fat infiltration in the lower leg. Red, green, yellow, and aqua colored regions represent anterior, lateral, deep, and superficial posterior compartments of the lower leg, respectively.

https://doi.org/10.1371/journal.pone.0310203.g005

Comparison of models using different learning methods and different amount of training data

Table 2 shows the results of pairwise comparison of the ADCs of few-supervised learning models using 2%, 5%, and 10% of the labeled data and semi-supervised learning models using 10% of the labeled data for training. Comparisons among few-supervised learning models using 2%, 5%, and 10% data sets revealed significantly improved segmentation performances as the amount of the labeled data for training increased, except for the comparison between models using central slice strategy with 5% and 10% data set for training. Pairwise comparisons between few-supervised and semi-supervised learning models using 10% of the labeled data showed significantly better segmentation performances of the semi-supervised models using slice central (Few- vs. Semi-supervised, 89.42±4.07% vs. 90.34±3.54%; p < 0.0001), lowermost slice strategy (Few- vs. Semi-supervised, 89.29±4.71% vs. 89.68±4.47%; p = 0.0081), three-slices strategy (Few- vs. Semi-supervised, 90.84±3.46% vs. 91.03±3.67%; p = 0.0293) and all-slices strategy (Few- vs Semi-supervised, 85.97±9.82% vs 89.71±4.23%; p < 0.0001). In contrast, the ADC of the few-supervised model using the uppermost slice strategy was significantly higher than that of the semi-supervised model (Few- vs Semi-supervised, 87.79±4.41% vs 91.03±3.67%; p = 0.0293). Further, the ADC of the fully-supervised model was 91.39±3.76%, which was significantly higher than that of the semi-supervised model using the three-slices strategy with 10% labeled set, which was the model that showed the highest segmentation performance (91.03±3.67%, p = 0.0246).

Download:

Table 2. Average dice coefficients of the few-supervised learning models using 2%, 5%, and 10% of the labeled data, and semi-supervised learning models using 10% of the labeled data for training.

https://doi.org/10.1371/journal.pone.0310203.t002

Discussion

Herein, we investigated the efficiency of several different training data labeling strategies and learning models to develop an automated segmentation model of the lower leg muscles in patients with CMT. Pairwise comparisons of ADCs were conducted for a few supervised learning models trained with 10% of labeled datasets using different labeling strategies. The results showed that the model using the three-slices strategy had segmentation performances significantly superior to those of the models using other labeling strategies. For the three-slices strategy, labeling was performed on the three axial image slices per subject: the uppermost, central, and lowermost image slices within the designated boundaries of the lower legs. The superior segmentation performance of the three-slices strategy compared with that of the single-slice strategies may be attributable to the fact that the three-slices strategy better accounts for interslice variation. Regardless of the specific level of images provided for training, a single image slice may not be sufficient to achieve optimal muscle segmentation performance in the lower legs, owing to the substantial differences in muscle compartment anatomy depending on the proximal-to-distal level [27]. The superior segmentation performance of the three-slices strategy compared with the all-slices strategy, which was trained with images from more variable proximal-to-distal levels and fewer subjects, might be due to the three-slices strategy’s better accounting for intersubject variation. It might therefore be necessary for the training dataset of the automated segmentation model to have a certain degree of inter-subject variation to obtain optimal segmentation performance by adjusting to the anatomical variation as well as the varying patterns and degrees of intramuscular fat infiltration of the lower leg muscles.

We evaluated the effect of unlabeled data utilization on the enhancement of muscle segmentation performance by comparing the performance of the models using few-supervised and semisupervised learning. Pairwise comparisons between the two models, for which we used the mean teacher model, showed significantly improved segmentation performance when semisupervised learning was used in most models. Leveraging unlabeled data in semi-supervised learning benefits the model in terms of learning the data distribution and improving the generalization and predictions of labeled data [35]. Training solely on labeled data may increase sensitivity to data noise, whereas leveraging unlabeled data enables the model to learn diverse patterns and variations, making it more robust to data noise [35]. With its potential to boost performance, the semisupervised model has shown promising results in image-segmentation tasks [33, 36]. Our results further suggest that the use of unlabeled data for training can substantially improve the network performance of lower leg muscle segmentation in patients with CMT. A semi-supervised model using the three-slices strategy with a 10% labeled set showed ADC of 91.03±3.67%, which we believe is comparable to the performance of the fully-supervised model (91.39±3.76%).

Our study has several limitations. First, the muscle segmentation models developed in our study were trained and tested on MRIs from a single institution, necessitating validation with MRIs from different acquisition parameters. We acknowledge that variability in MRI protocols across institutions may influence the generalizability of the model. Second, models were developed using two-dimensional T1-weighted spin-echo images. A three-dimensional gradient echo Dixon-based MRI sequence with a larger number of axial images and thinner slices may be more valuable for automated muscle segmentation [3, 4]. However, using such a sequence would require significantly more time and effort in the ground truth segmentation process because of the higher number of axial images per designated anatomical region compared to two-dimensional T1-weighted spin-echo images. We believe that our results can guide the design of a strategy for ground-truth segmentation using a three-dimensional MRI sequence. Third, we did not evaluate more advanced deep learning models, such as the nnU-net or residual U-Net [37, 38], which might have the potential to enhance the segmentation performance of the lower leg muscles in patients with CMT. Finally, potential bias may be present due to manual segmentations being performed by a single rater.

In conclusion, we identified that performing ground truth segmentation using a three-slice-based strategy, with labeling of the uppermost, central, and lowermost image slices of a subset of subjects within the designated anatomical volume of the lower legs in patients with CMT, was the most efficient strategy for building an automated muscle segmentation of axial T1-weighted MRI. In addition, our results indicate that a semi-supervised model using unlabeled data with a mean teacher model significantly improves the performance of the model. Our findings may enhance the efficiency of the ground truth segmentation process in future studies on automated muscle segmentation, addressing a significant challenge in developing an automated segmentation model. Based on the conclusions drawn about the most efficient method for ground truth segmentation in this study, we aim to develop an automated segmentation model of axial three-dimensional gradient echo Dixon-based MRI sequences in the lower legs of patients with CMT in future studies.

Supporting information

S1 File. 2D U-Net for segmentation of the four muscle compartments on MRI.

https://doi.org/10.1371/journal.pone.0310203.s001

(DOCX)

S2 File. Raw data of the segmentation models.

https://doi.org/10.1371/journal.pone.0310203.s002

(PDF)

References

1. Pareyson D, Marchesi C. Diagnosis, natural history, and management of Charcot-Marie-Tooth disease. Lancet Neurol. 2009; 8: 654–667. pmid:19539237
- View Article
- PubMed/NCBI
- Google Scholar
2. Bas J, Ogier AC, Le Troter A, Delmont E, Leporq B, Pini L, et al. Fat fraction distribution in lower limb muscles of patients with CMT1A: A quantitative MRI study. Neurology. 2020; 94: e1480–e1487. pmid:31980579
- View Article
- PubMed/NCBI
- Google Scholar
3. Kim HS, Yoon YC, Choi BO, Jin W, Cha JG. Muscle fat quantification using magnetic resonance imaging: case-control study of Charcot-Marie-Tooth disease patients and volunteers. J Cachexia Sarcopenia Muscle. 2019; 10: 574–585. pmid:30873759
- View Article
- PubMed/NCBI
- Google Scholar
4. Morrow JM, Evans MRB, Grider T, Sinclair CDJ, Thedens D, Shah S, et al. Validation of MRC Centre MRI calf muscle fat fraction protocol as an outcome measure in CMT1A. Neurology. 2018; 91: e1125–e1129. pmid:30120135
- View Article
- PubMed/NCBI
- Google Scholar
5. Morrow JM, Sinclair CD, Fischmann A, Machado PM, Reilly MM, Yousry TA, et al. MRI biomarker assessment of neuromuscular disease progression: a prospective observational cohort study. Lancet Neurol. 2016; 15: 65–77. pmid:26549782
- View Article
- PubMed/NCBI
- Google Scholar
6. Kanber B, Morrow JM, Klickovic U, Wastling S, Shah S, Fratta P, et al. Musclesense: a Trained, Artificial Neural Network for the Anatomical Segmentation of Lower Limb Magnetic Resonance Images in Neuromuscular Diseases. Neuroinformatics. 2021; 19: 379–383. pmid:32892313
- View Article
- PubMed/NCBI
- Google Scholar
7. Ogier AC, Hostin MA, Bellemare ME, Bendahan D. Overview of MR Image Segmentation Strategies in Neuromuscular Disorders. Front Neurol. 2021; 12: 625308. pmid:33841299
- View Article
- PubMed/NCBI
- Google Scholar
8. Cornett KMD, Wojciechowski E, Sman AD, Walker T, Menezes MP, Bray P, et al. Magnetic resonance imaging of the anterior compartment of the lower leg is a biomarker for weakness, disability, and impaired gait in childhood Charcot-Marie-Tooth disease. Muscle Nerve. 2019; 59: 213–217. pmid:30265406
- View Article
- PubMed/NCBI
- Google Scholar
9. Kim YJ, Kim HS, Lee JH, Yoon YC, Choi BO. Magnetic resonance imaging-based lower limb muscle evaluation in Charcot-Marie-Tooth disease type 1A patients and its correlation with clinical data. Sci Rep. 2022; 12: 16622. pmid:36198750
- View Article
- PubMed/NCBI
- Google Scholar
10. Stilwell G, Kilcoyne RF, Sherman JL. Patterns of muscle atrophy in the lower limbs in patients with Charcot-Marie-Tooth disease as measured by magnetic resonance imaging. J Foot Ankle Surg. 1995; 34: 583–586; discussion 596. pmid:8646212
- View Article
- PubMed/NCBI
- Google Scholar
11. Gallardo E, Claeys KG, Nelis E, Garcia A, Canga A, Combarros O, et al. Magnetic resonance imaging findings of leg musculature in Charcot-Marie-Tooth disease type 2 due to dynamin 2 mutation. J Neurol. 2008; 255: 986–992. pmid:18560793
- View Article
- PubMed/NCBI
- Google Scholar
12. Gaeta M, Mileto A, Mazzeo A, Minutoli F, Di Leo R, Settineri N, et al. MRI findings, patterns of disease distribution, and muscle fat fraction calculation in five patients with Charcot-Marie-Tooth type 2 F disease. Skeletal Radiol. 2012; 41: 515–524. pmid:21611841
- View Article
- PubMed/NCBI
- Google Scholar
13. Ding J, Cao P, Chang HC, Gao Y, Chan SHS, Vardhanabhuti V. Deep learning-based thigh muscle segmentation for reproducible fat fraction quantification using fat-water decomposition MRI. Insights Imaging. 2020; 11: 128. pmid:33252711
- View Article
- PubMed/NCBI
- Google Scholar
14. Gadermayr M, Disch C, Muller M, Merhof D, Gess B. A comprehensive study on automated muscle segmentation for assessing fat infiltration in neuromuscular diseases. Magn Reson Imaging. 2018; 48: 20–26. pmid:29269318
- View Article
- PubMed/NCBI
- Google Scholar
15. Goutallier D, Postel JM, Bernageau J, Lavau L, Voisin MC. Fatty muscle degeneration in cuff ruptures. Pre- and postoperative evaluation by CT scan. Clin Orthop Relat Res. 1994: 78–83. pmid:8020238
- View Article
- PubMed/NCBI
- Google Scholar
16. Rong Y, Xiang D, Zhu W, Shi F, Gao E, Fan Z, et al. Deriving external forces via convolutional neural networks for biomedical image segmentation. Biomed Opt Express. 2019; 10: 3800–3814. pmid:31452976
- View Article
- PubMed/NCBI
- Google Scholar
17. Sudre CH, Li W, Vercauteren T, Ourselin S, Jorge Cardoso M. Generalised Dice Overlap as a Deep Learning Loss Function for Highly Unbalanced Segmentations. Deep Learn Med Image Anal Multimodal Learn Clin Decis Support (2017). 2017; 2017: 240–248. pmid:34104926
- View Article
- PubMed/NCBI
- Google Scholar
18. Zhou Z, Shin J, Zhang L, Gurudu S, Gotway M, Liang J. Fine-tuning Convolutional Neural Networks for Biomedical Image Analysis: Actively and Incrementally. Proc IEEE Comput Soc Conf Comput Vis Pattern Recognit. 2017; 2017: 4761–4772. pmid:30337799
- View Article
- PubMed/NCBI
- Google Scholar
19. Attarian S, Young P, Brannagan TH, Adams D, Van Damme P, Thomas FP, et al. A double-blind, placebo-controlled, randomized trial of PXT3003 for the treatment of Charcot-Marie-Tooth type 1A. Orphanet J Rare Dis. 2021; 16: 433. pmid:34656144
- View Article
- PubMed/NCBI
- Google Scholar
20. Klein D, Patzko A, Schreiber D, van Hauwermeiren A, Baier M, Groh J, et al. Targeting the colony stimulating factor 1 receptor alleviates two forms of Charcot-Marie-Tooth disease in mice. Brain. 2015; 138: 3193–3205. pmid:26297559
- View Article
- PubMed/NCBI
- Google Scholar
21. Pisciotta C, Pareyson D. Gene therapy and other novel treatment approaches for Charcot-Marie-Tooth disease. Neuromuscul Disord. 2023; 33: 627–635. pmid:37455204
- View Article
- PubMed/NCBI
- Google Scholar
22. Kemnitz J, Baumgartner CF, Eckstein F, Chaudhari A, Ruhdorfer A, Wirth W, et al. Clinical evaluation of fully automated thigh muscle and adipose tissue segmentation using a U-Net deep learning architecture in context of osteoarthritic knee pain. MAGMA. 2020; 33: 483–493. pmid:31872357
- View Article
- PubMed/NCBI
- Google Scholar
23. Mesbah S, Shalaby AM, Stills S, Soliman AM, Willhite A, Harkema SJ, et al. Novel stochastic framework for automatic segmentation of human thigh MRI volumes and its applications in spinal cord injured individuals. PLoS One. 2019; 14: e0216487. pmid:31071158
- View Article
- PubMed/NCBI
- Google Scholar
24. Kim HS, Kim HJ, Nam SH, Kim SB, Choi YJ, Lee KS, et al. Clinical and Neuroimaging Features in Charcot-Marie-Tooth Patients with GDAP1 Mutations. J Clin Neurol. 2021; 17: 52–62. pmid:33480199
- View Article
- PubMed/NCBI
- Google Scholar
25. Kwon HM, Kim HS, Kim SB, Park JH, Nam DE, Lee AJ, et al. Clinical and Neuroimaging Features in Charcot-Marie-Tooth Patients with GNB4 Mutations. Life (Basel). 2021; 11 pmid:34071515
- View Article
- PubMed/NCBI
- Google Scholar
26. Park J, Kim HS, Kwon HM, Kim J, Nam SH, Jung NY, et al. Identification and clinical characterization of Charcot-Marie-Tooth disease type 1C patients with LITAF p.G112S mutation. Genes Genomics. 2022; 44: 1007–1016. pmid:35608774
- View Article
- PubMed/NCBI
- Google Scholar
27. Anderson MW, Temple HT, Dussault RG, Kaplan PA. Compartmental anatomy: relevance to staging and biopsy of musculoskeletal tumors. AJR Am J Roentgenol. 1999; 173: 1663–1671. pmid:10584817
- View Article
- PubMed/NCBI
- Google Scholar
28. Bosma JS, Saha A, Hosseinzadeh M, Slootweg I, de Rooij M, Huisman H. Semisupervised Learning with Report-guided Pseudo Labels for Deep Learning-based Prostate Cancer Detection Using Biparametric MRI. Radiol Artif Intell. 2023; 5: e230031. pmid:37795142
- View Article
- PubMed/NCBI
- Google Scholar
29. Ge C, Gu IY, Jakola AS, Yang J. Deep semi-supervised learning for brain tumor classification. BMC Med Imaging. 2020; 20: 87. pmid:32727476
- View Article
- PubMed/NCBI
- Google Scholar
30. Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation. In: Navab N, Hornegger J, Wells WM, Frangi AF, editors. Medical Image Computing and Computer-Assisted Intervention: MICCAI 2015. Lecture Notes in Computer Science: Springer International Publishing; 2015. pp. 234–241.
31. Khouy M, Jabrane Y, Ameur M, Hajjam El Hassani A. Medical Image Segmentation Using Automatic Optimized U-Net Architecture Based on Genetic Algorithm. J Pers Med. 2023; 13: 1298. pmid:37763066
- View Article
- PubMed/NCBI
- Google Scholar
32. Tarvainen A, Valpola H. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. 2017 [Cited 2023 December 1]. Available from: https://arxiv.org/abs/1703.01780.
- View Article
- Google Scholar
33. Huo J, Ouyang X, Si L, Xuan K, Wang S, Yao W, et al. Automatic Grading Assessments for Knee MRI Cartilage Defects via Self-ensembling Semi-supervised Learning with Dual-Consistency. Med Image Anal. 2022; 80: 102508. pmid:35759870
- View Article
- PubMed/NCBI
- Google Scholar
34. Laine S, Aila T. Temporal ensembling for semi-supervised learning 2016 [Cited 2023 December 1]. Available from: https://arxiv.org/abs/1610.02242.
- View Article
- Google Scholar
35. Mey A, Loog M. Improved Generalization in Semi-Supervised Learning: A Survey of Theoretical Results. IEEE Trans Pattern Anal Mach Intell. 2023; 45: 4747–4767. pmid:35984799
- View Article
- PubMed/NCBI
- Google Scholar
36. Decourt C, Duong L. Semi-supervised generative adversarial networks for the segmentation of the left ventricle in pediatric MRI. Comput Biol Med. 2020; 123: 103884. pmid:32658792
- View Article
- PubMed/NCBI
- Google Scholar
37. Peng Y, Xu Y, Wang M, Zhang H, Xie J. The nnU-Net based method for automatic segmenting fetal brain tissues. Health Inf Sci Syst. 2023; 11: 17. pmid:36998806
- View Article
- PubMed/NCBI
- Google Scholar
38. Li P, Li Z, Wang Z, Li C, Wang M. mResU-Net: multi-scale residual U-Net-based brain tumor segmentation from multimodal MRI. Med Biol Eng Comput. 2023 pmid:37981627
- View Article
- PubMed/NCBI
- Google Scholar

[ref1] 1. Pareyson D, Marchesi C. Diagnosis, natural history, and management of Charcot-Marie-Tooth disease. Lancet Neurol. 2009; 8: 654–667. pmid:19539237
View Article
PubMed/NCBI
Google Scholar

[2] View Article

[3] PubMed/NCBI

[4] Google Scholar

[ref2] 2. Bas J, Ogier AC, Le Troter A, Delmont E, Leporq B, Pini L, et al. Fat fraction distribution in lower limb muscles of patients with CMT1A: A quantitative MRI study. Neurology. 2020; 94: e1480–e1487. pmid:31980579
View Article
PubMed/NCBI
Google Scholar

[6] View Article

[7] PubMed/NCBI

[8] Google Scholar

[ref3] 3. Kim HS, Yoon YC, Choi BO, Jin W, Cha JG. Muscle fat quantification using magnetic resonance imaging: case-control study of Charcot-Marie-Tooth disease patients and volunteers. J Cachexia Sarcopenia Muscle. 2019; 10: 574–585. pmid:30873759
View Article
PubMed/NCBI
Google Scholar

[10] View Article

[11] PubMed/NCBI

[12] Google Scholar

[ref4] 4. Morrow JM, Evans MRB, Grider T, Sinclair CDJ, Thedens D, Shah S, et al. Validation of MRC Centre MRI calf muscle fat fraction protocol as an outcome measure in CMT1A. Neurology. 2018; 91: e1125–e1129. pmid:30120135
View Article
PubMed/NCBI
Google Scholar

[14] View Article

[15] PubMed/NCBI

[16] Google Scholar

[ref5] 5. Morrow JM, Sinclair CD, Fischmann A, Machado PM, Reilly MM, Yousry TA, et al. MRI biomarker assessment of neuromuscular disease progression: a prospective observational cohort study. Lancet Neurol. 2016; 15: 65–77. pmid:26549782
View Article
PubMed/NCBI
Google Scholar

[18] View Article

[19] PubMed/NCBI

[20] Google Scholar

[ref6] 6. Kanber B, Morrow JM, Klickovic U, Wastling S, Shah S, Fratta P, et al. Musclesense: a Trained, Artificial Neural Network for the Anatomical Segmentation of Lower Limb Magnetic Resonance Images in Neuromuscular Diseases. Neuroinformatics. 2021; 19: 379–383. pmid:32892313
View Article
PubMed/NCBI
Google Scholar

[22] View Article

[23] PubMed/NCBI

[24] Google Scholar

[ref7] 7. Ogier AC, Hostin MA, Bellemare ME, Bendahan D. Overview of MR Image Segmentation Strategies in Neuromuscular Disorders. Front Neurol. 2021; 12: 625308. pmid:33841299
View Article
PubMed/NCBI
Google Scholar

[26] View Article

[27] PubMed/NCBI

[28] Google Scholar

[ref8] 8. Cornett KMD, Wojciechowski E, Sman AD, Walker T, Menezes MP, Bray P, et al. Magnetic resonance imaging of the anterior compartment of the lower leg is a biomarker for weakness, disability, and impaired gait in childhood Charcot-Marie-Tooth disease. Muscle Nerve. 2019; 59: 213–217. pmid:30265406
View Article
PubMed/NCBI
Google Scholar

[30] View Article

[31] PubMed/NCBI

[32] Google Scholar

[ref9] 9. Kim YJ, Kim HS, Lee JH, Yoon YC, Choi BO. Magnetic resonance imaging-based lower limb muscle evaluation in Charcot-Marie-Tooth disease type 1A patients and its correlation with clinical data. Sci Rep. 2022; 12: 16622. pmid:36198750
View Article
PubMed/NCBI
Google Scholar

[34] View Article

[35] PubMed/NCBI

[36] Google Scholar

[ref10] 10. Stilwell G, Kilcoyne RF, Sherman JL. Patterns of muscle atrophy in the lower limbs in patients with Charcot-Marie-Tooth disease as measured by magnetic resonance imaging. J Foot Ankle Surg. 1995; 34: 583–586; discussion 596. pmid:8646212
View Article
PubMed/NCBI
Google Scholar

[38] View Article

[39] PubMed/NCBI

[40] Google Scholar

[ref11] 11. Gallardo E, Claeys KG, Nelis E, Garcia A, Canga A, Combarros O, et al. Magnetic resonance imaging findings of leg musculature in Charcot-Marie-Tooth disease type 2 due to dynamin 2 mutation. J Neurol. 2008; 255: 986–992. pmid:18560793
View Article
PubMed/NCBI
Google Scholar

[42] View Article

[43] PubMed/NCBI

[44] Google Scholar

[ref12] 12. Gaeta M, Mileto A, Mazzeo A, Minutoli F, Di Leo R, Settineri N, et al. MRI findings, patterns of disease distribution, and muscle fat fraction calculation in five patients with Charcot-Marie-Tooth type 2 F disease. Skeletal Radiol. 2012; 41: 515–524. pmid:21611841
View Article
PubMed/NCBI
Google Scholar

[46] View Article

[47] PubMed/NCBI

[48] Google Scholar

[ref13] 13. Ding J, Cao P, Chang HC, Gao Y, Chan SHS, Vardhanabhuti V. Deep learning-based thigh muscle segmentation for reproducible fat fraction quantification using fat-water decomposition MRI. Insights Imaging. 2020; 11: 128. pmid:33252711
View Article
PubMed/NCBI
Google Scholar

[50] View Article

[51] PubMed/NCBI

[52] Google Scholar

[ref14] 14. Gadermayr M, Disch C, Muller M, Merhof D, Gess B. A comprehensive study on automated muscle segmentation for assessing fat infiltration in neuromuscular diseases. Magn Reson Imaging. 2018; 48: 20–26. pmid:29269318
View Article
PubMed/NCBI
Google Scholar

[54] View Article

[55] PubMed/NCBI

[56] Google Scholar

[ref15] 15. Goutallier D, Postel JM, Bernageau J, Lavau L, Voisin MC. Fatty muscle degeneration in cuff ruptures. Pre- and postoperative evaluation by CT scan. Clin Orthop Relat Res. 1994: 78–83. pmid:8020238
View Article
PubMed/NCBI
Google Scholar

[58] View Article

[59] PubMed/NCBI

[60] Google Scholar

[ref16] 16. Rong Y, Xiang D, Zhu W, Shi F, Gao E, Fan Z, et al. Deriving external forces via convolutional neural networks for biomedical image segmentation. Biomed Opt Express. 2019; 10: 3800–3814. pmid:31452976
View Article
PubMed/NCBI
Google Scholar

[62] View Article

[63] PubMed/NCBI

[64] Google Scholar

[ref17] 17. Sudre CH, Li W, Vercauteren T, Ourselin S, Jorge Cardoso M. Generalised Dice Overlap as a Deep Learning Loss Function for Highly Unbalanced Segmentations. Deep Learn Med Image Anal Multimodal Learn Clin Decis Support (2017). 2017; 2017: 240–248. pmid:34104926
View Article
PubMed/NCBI
Google Scholar

[66] View Article

[67] PubMed/NCBI

[68] Google Scholar

[ref18] 18. Zhou Z, Shin J, Zhang L, Gurudu S, Gotway M, Liang J. Fine-tuning Convolutional Neural Networks for Biomedical Image Analysis: Actively and Incrementally. Proc IEEE Comput Soc Conf Comput Vis Pattern Recognit. 2017; 2017: 4761–4772. pmid:30337799
View Article
PubMed/NCBI
Google Scholar

[70] View Article

[71] PubMed/NCBI

[72] Google Scholar

[ref19] 19. Attarian S, Young P, Brannagan TH, Adams D, Van Damme P, Thomas FP, et al. A double-blind, placebo-controlled, randomized trial of PXT3003 for the treatment of Charcot-Marie-Tooth type 1A. Orphanet J Rare Dis. 2021; 16: 433. pmid:34656144
View Article
PubMed/NCBI
Google Scholar

[74] View Article

[75] PubMed/NCBI

[76] Google Scholar

[ref20] 20. Klein D, Patzko A, Schreiber D, van Hauwermeiren A, Baier M, Groh J, et al. Targeting the colony stimulating factor 1 receptor alleviates two forms of Charcot-Marie-Tooth disease in mice. Brain. 2015; 138: 3193–3205. pmid:26297559
View Article
PubMed/NCBI
Google Scholar

[78] View Article

[79] PubMed/NCBI

[80] Google Scholar

[ref21] 21. Pisciotta C, Pareyson D. Gene therapy and other novel treatment approaches for Charcot-Marie-Tooth disease. Neuromuscul Disord. 2023; 33: 627–635. pmid:37455204
View Article
PubMed/NCBI
Google Scholar

[82] View Article

[83] PubMed/NCBI

[84] Google Scholar

[ref22] 22. Kemnitz J, Baumgartner CF, Eckstein F, Chaudhari A, Ruhdorfer A, Wirth W, et al. Clinical evaluation of fully automated thigh muscle and adipose tissue segmentation using a U-Net deep learning architecture in context of osteoarthritic knee pain. MAGMA. 2020; 33: 483–493. pmid:31872357
View Article
PubMed/NCBI
Google Scholar

[86] View Article

[87] PubMed/NCBI

[88] Google Scholar

[ref23] 23. Mesbah S, Shalaby AM, Stills S, Soliman AM, Willhite A, Harkema SJ, et al. Novel stochastic framework for automatic segmentation of human thigh MRI volumes and its applications in spinal cord injured individuals. PLoS One. 2019; 14: e0216487. pmid:31071158
View Article
PubMed/NCBI
Google Scholar

[90] View Article

[91] PubMed/NCBI

[92] Google Scholar

[ref24] 24. Kim HS, Kim HJ, Nam SH, Kim SB, Choi YJ, Lee KS, et al. Clinical and Neuroimaging Features in Charcot-Marie-Tooth Patients with GDAP1 Mutations. J Clin Neurol. 2021; 17: 52–62. pmid:33480199
View Article
PubMed/NCBI
Google Scholar

[94] View Article

[95] PubMed/NCBI

[96] Google Scholar

[ref25] 25. Kwon HM, Kim HS, Kim SB, Park JH, Nam DE, Lee AJ, et al. Clinical and Neuroimaging Features in Charcot-Marie-Tooth Patients with GNB4 Mutations. Life (Basel). 2021; 11 pmid:34071515
View Article
PubMed/NCBI
Google Scholar

[98] View Article

[99] PubMed/NCBI

[100] Google Scholar

[ref26] 26. Park J, Kim HS, Kwon HM, Kim J, Nam SH, Jung NY, et al. Identification and clinical characterization of Charcot-Marie-Tooth disease type 1C patients with LITAF p.G112S mutation. Genes Genomics. 2022; 44: 1007–1016. pmid:35608774
View Article
PubMed/NCBI
Google Scholar

[102] View Article

[103] PubMed/NCBI

[104] Google Scholar

[ref27] 27. Anderson MW, Temple HT, Dussault RG, Kaplan PA. Compartmental anatomy: relevance to staging and biopsy of musculoskeletal tumors. AJR Am J Roentgenol. 1999; 173: 1663–1671. pmid:10584817
View Article
PubMed/NCBI
Google Scholar

[106] View Article

[107] PubMed/NCBI

[108] Google Scholar

[ref28] 28. Bosma JS, Saha A, Hosseinzadeh M, Slootweg I, de Rooij M, Huisman H. Semisupervised Learning with Report-guided Pseudo Labels for Deep Learning-based Prostate Cancer Detection Using Biparametric MRI. Radiol Artif Intell. 2023; 5: e230031. pmid:37795142
View Article
PubMed/NCBI
Google Scholar

[110] View Article

[111] PubMed/NCBI

[112] Google Scholar

[ref29] 29. Ge C, Gu IY, Jakola AS, Yang J. Deep semi-supervised learning for brain tumor classification. BMC Med Imaging. 2020; 20: 87. pmid:32727476
View Article
PubMed/NCBI
Google Scholar

[114] View Article

[115] PubMed/NCBI

[116] Google Scholar

[ref30] 30. Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation. In: Navab N, Hornegger J, Wells WM, Frangi AF, editors. Medical Image Computing and Computer-Assisted Intervention: MICCAI 2015. Lecture Notes in Computer Science: Springer International Publishing; 2015. pp. 234–241.

[ref31] 31. Khouy M, Jabrane Y, Ameur M, Hajjam El Hassani A. Medical Image Segmentation Using Automatic Optimized U-Net Architecture Based on Genetic Algorithm. J Pers Med. 2023; 13: 1298. pmid:37763066
View Article
PubMed/NCBI
Google Scholar

[119] View Article

[120] PubMed/NCBI

[121] Google Scholar

[ref32] 32. Tarvainen A, Valpola H. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. 2017 [Cited 2023 December 1]. Available from: https://arxiv.org/abs/1703.01780.
View Article
Google Scholar

[123] View Article

[124] Google Scholar

[ref33] 33. Huo J, Ouyang X, Si L, Xuan K, Wang S, Yao W, et al. Automatic Grading Assessments for Knee MRI Cartilage Defects via Self-ensembling Semi-supervised Learning with Dual-Consistency. Med Image Anal. 2022; 80: 102508. pmid:35759870
View Article
PubMed/NCBI
Google Scholar

[126] View Article

[127] PubMed/NCBI

[128] Google Scholar

[ref34] 34. Laine S, Aila T. Temporal ensembling for semi-supervised learning 2016 [Cited 2023 December 1]. Available from: https://arxiv.org/abs/1610.02242.
View Article
Google Scholar

[130] View Article

[131] Google Scholar

[ref35] 35. Mey A, Loog M. Improved Generalization in Semi-Supervised Learning: A Survey of Theoretical Results. IEEE Trans Pattern Anal Mach Intell. 2023; 45: 4747–4767. pmid:35984799
View Article
PubMed/NCBI
Google Scholar

[133] View Article

[134] PubMed/NCBI

[135] Google Scholar

[ref36] 36. Decourt C, Duong L. Semi-supervised generative adversarial networks for the segmentation of the left ventricle in pediatric MRI. Comput Biol Med. 2020; 123: 103884. pmid:32658792
View Article
PubMed/NCBI
Google Scholar

[137] View Article

[138] PubMed/NCBI

[139] Google Scholar

[ref37] 37. Peng Y, Xu Y, Wang M, Zhang H, Xie J. The nnU-Net based method for automatic segmenting fetal brain tissues. Health Inf Sci Syst. 2023; 11: 17. pmid:36998806
View Article
PubMed/NCBI
Google Scholar

[141] View Article

[142] PubMed/NCBI

[143] Google Scholar

[ref38] 38. Li P, Li Z, Wang Z, Li C, Wang M. mResU-Net: multi-scale residual U-Net-based brain tumor segmentation from multimodal MRI. Med Biol Eng Comput. 2023 pmid:37981627
View Article
PubMed/NCBI
Google Scholar

[145] View Article

[146] PubMed/NCBI

[147] Google Scholar

Figures

Abstract

Introduction

Materials and methods

Patient selection

MRI protocol

Ground truth segmentation

Comparisons of models using different training data labeling strategies

Comparison of models using different learning methods and different amount of training data

Image preprocessing

2D U-net: 2D segmentation model

Semi-supervised deep learning model

Implementation details

Statistical analyses

Results

Comparisons of models using different training data labeling strategies

Comparison of models using different learning methods and different amount of training data

Discussion

Supporting information

S1 File. 2D U-Net for segmentation of the four muscle compartments on MRI.

S2 File. Raw data of the segmentation models.

References