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

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.

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”.

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.

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

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.

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.